CN103473759B

CN103473759B - The twilight image method for extracting remarkable configuration that a kind of WKPCA homogeneity degree correction nCRF suppresses

Info

Publication number: CN103473759B
Application number: CN201310253229.1A
Authority: CN
Inventors: 柏连发; 张毅; 陈钱; 顾国华; 韩静; 岳江; 祁伟; 金左轮
Original assignee: Nanjing University of Science and Technology
Current assignee: Nanjing University of Science and Technology
Priority date: 2013-06-24
Filing date: 2013-06-24
Publication date: 2016-07-06
Anticipated expiration: 2033-06-24
Also published as: CN103473759A

Abstract

The invention discloses a method for extracting salient contours of low-light images suppressed by WKPCA homogeneity correction nCRF. First, a WKPCA algorithm is proposed, which performs feature vector angle matching (FAM) weighting on each feature vector in the high-dimensional feature space, weakens or eliminates the interference of pathological or abnormal feature data in the CRF region, and extracts the principal components of the CRF more accurately; here Based on this, a concept and calculation method of homogeneity is defined, and the environment-center homogeneity is calculated through the projection of nCRF eigenvectors on the central principal component; finally, based on the homogeneity, each suppression amount in nCRF is corrected, so that the homogeneity The amount of mutual inhibition in the homogeneous area is large, the amount of inhibition in the heterogeneous area is small or no mutual inhibition, and at the same time, the self-inhibition of the contour elements is weakened as much as possible, so as to improve the accuracy of the inhibition. Therefore, the model proposed by the present invention can detect environment-center differences more comprehensively, reduce noise interference, suppress texture details more accurately, and improve contour response strength and integrity.

Description

A salient contour extraction method for low-light images based on WKPCA homogeneity correction nCRF suppression

技术领域technical field

本发明属于一种基于视觉建模的复杂场景下微光图像显著轮廓提取方法，特别是一种WKPCA同质度校正nCRF抑制的微光图像显著轮廓提取方法。The invention belongs to a method for extracting salient contours of low-light images in complex scenes based on visual modeling, in particular to a method for extracting salient contours of low-light images suppressed by WKPCA homogeneity correction nCRF.

背景技术Background technique

轮廓提取在夜视图像理解分析方面发挥着重要作用。目前夜视目标探测识别方面的应用大部分是针对自然场景的，因此微光图像中包含了大量的自然纹理（例如树和草），传统的边缘检测算子的作用结果保留大量非轮廓的边缘成分（canny算子）。而且微光图像本身噪声干扰强，如何针对微光图像特征，去除这些由纹理和噪声所产生的局部非兴趣边缘，并且保持轮廓的完整性是夜视图像轮廓检测主要面临的问题。Contour extraction plays an important role in understanding and analyzing night vision images. At present, most of the applications of night vision target detection and recognition are aimed at natural scenes, so low-light images contain a large number of natural textures (such as trees and grass), and the results of traditional edge detection operators retain a large number of non-contour edges. Composition (canny operator). Moreover, the low-light image itself has strong noise interference. How to remove these local uninteresting edges generated by texture and noise and maintain the integrity of the contour is the main problem for night vision image contour detection.

针对复杂场景的轮廓提取问题提出了诸多解决方法，其中基于生物视觉机理的非经典感受野模型的轮廓提取在高质量可见光图像中获得了显著效果。视皮层（V1）神经元感受野（CRF）的大外周（非经典感受野nCRF）对CRF起调制作用，这种调制主要是抑制性的，能够实现同质区域相互抑制，而使得孤立的边缘要比群体边缘更为显著。基于侧抑制区的仿生模型，较好地去除了背景纹理产生的边缘。Grigorescu等人（ContourdetectionbasedonNonclassicalReceptiveFieldinhibition;ImprovedContourDetectionbyNon-classicalReceptiveFieldInhibition;Contourandboundarydetectionimprovedbysurroundsuppressionoftextureedges）利用nCRF的抑制特性进行轮廓检测，利用环境对中心的方向抑制，减少了环境纹理的影响，并提出各向异性抑制和各向同性抑制模型。Giuseppe等人（ABiologicallyMotivatedMultiresolutionApproachtoContourDetection）基于贝叶斯降噪和环境抑制技术提出一种生物启发的多分辨率轮廓检测技术。Ursino等人（Amodelofcontourextractionincludingmultiplescales,flexibleinhibitionandattention）引入层级注意机制，考虑不同尺度下背景抑制的轮廓提取。这些抑制模型能够一定程度的抑制背景纹理，提取出显著轮廓；但没有对中心环境进行特征差异分析，不能很好的解决异质成份互抑制和轮廓元素自抑制问题，可能导致轮廓响应微弱和轮廓间断。Many solutions have been proposed for the contour extraction of complex scenes. Among them, the contour extraction of the non-classical receptive field model based on the mechanism of biological vision has achieved remarkable results in high-quality visible light images. The large periphery of the receptive field (CRF) of neurons in the visual cortex (V1) (non-classical receptive field nCRF) modulates the CRF, and this modulation is mainly inhibitory, enabling mutual inhibition of homogeneous regions while making isolated edges are more significant than group margins. Based on the bionic model of the side suppression zone, the edge generated by the background texture is better removed. Grigorescu et al. (Contour detection based on Nonclassical Receptive Field inhibition; Improved Contour Detection by Non-classical Receptive Field Inhibition; Contour and boundary detection improved by surround suppression of textureedges) use the suppression characteristics of nCRF for contour detection, use the environment's direction suppression to the center, reduce the influence of environmental texture, and propose an anisotropic suppression model and anisotropic suppression. Giuseppe et al. (ABiologicallyMotivatedMultiresolutionApproachtoContourDetection) proposed a biologically inspired multiresolution contour detection technique based on Bayesian noise reduction and environmental suppression techniques. Ursino et al. (Amodel of contour extraction including multiple scales, flexible inhibition and attention) introduce a hierarchical attention mechanism to consider contour extraction for background suppression at different scales. These suppression models can suppress the background texture to a certain extent and extract significant contours; however, they do not perform feature difference analysis on the central environment, and cannot solve the problems of mutual suppression of heterogeneous components and self-suppression of contour elements, which may lead to weak contour responses and Intermittent.

为减小共线轮廓抑制的作用，桑农等人（Contourdetectionbasedoninhibitionofprimaryvisualcortex）根据nCRF刺激方位与CRF刺激方位差异，对抑制作用加权，建立了基于侧抑制区的蝶形模型。Zeng等人（Center-surroundinteractionwithadaptiveinhibition:Acomputationalmodelforcontourdetection;Contourdetectionbasedonanon-classicalreceptivefieldmodelwithbutterfly-shapedinhibitionsubregions）提出一种双尺度轮廓提取方法以及一种改进的方向选择抑制模型，该模型采用蝶形区域计算环境抑制。Giuseppe等人（Animprovedmodelforsurroundsuppressionbysteerablefiltersandmultilevelinhibitionwithapplicationtocontourdetection）利用可变滤波器和多层级抑制提出一种环境抑制模型。方向差异加权的抑制模型一定程度上减少了轮廓自抑制作用，但存在异向异质区域相互抑制的情况。而且相对于高质量的可见光图像，微光图像噪声干扰严重，轮廓局部方向特征不突出，环境抑制不准确。一方面高噪声、低对比度导致微光图像背景纹理无法完全抑制；另一方面抑制作用削弱轮廓强度，轮廓受到周边纹理的抑制，容易出现断裂，影响后续目标识别。In order to reduce the effect of collinear contour inhibition, Sannon et al. (Contour detection based on inhibition of primary visual cortex) weighted the inhibitory effect according to the difference between nCRF stimulation orientation and CRF stimulation orientation, and established a butterfly model based on the lateral inhibition area. Zeng et al. (Center-surround interaction with adaptive inhibition: Acomputational model for contour detection; Contour detection based on ananon-classical receptive field model with butterfly-shaped inhibition subregions) proposed a dual-scale contour extraction method and an improved direction-selective inhibition model that uses butterfly-shaped regions to compute environmental suppression. Giuseppe et al. (Animprovedmodelforsurroundsuppressionbysteerablefiltersandmultilevelinhibitionwithapplicationtocontourdetection) proposed an environmental suppression model using variable filters and multilevel suppression. The inhibition model weighted by direction difference reduces the self-inhibition effect of the contour to a certain extent, but there is a situation that the heterogeneous regions inhibit each other. Moreover, compared with high-quality visible light images, low-light images suffer from severe noise interference, local directional features of contours are not prominent, and environmental suppression is inaccurate. On the one hand, high noise and low contrast make the background texture of the low-light image unable to be completely suppressed; on the other hand, the suppression effect weakens the contour strength, and the contour is suppressed by the surrounding texture, which is prone to breakage and affects subsequent target recognition.

发明内容Contents of the invention

本发明的目的在于提供WKPCA同质度校正nCRF抑制的微光图像显著轮廓提取方法，该方法能够解决单一方位差加权抑制的不足，结合生物视觉机理，针对微光图像多维特征差异，从复杂场景下微光图像中有效抑制噪声和纹理，提取显著轮廓。The purpose of the present invention is to provide a method for extracting salient contours of low-light images suppressed by WKPCA homogeneity correction nCRF, which can solve the deficiency of single azimuth weighted suppression, combined with the mechanism of biological vision, aiming at multi-dimensional feature differences of low-light images, from complex scenes Effectively suppress noise and texture in low-light images, and extract salient contours.

实现本发明目的的技术解决方案为：首先在高维特征空间研究一种特征向量角匹配(FAMFeaturevectorAngleMatching)加权KPCA(WKPCAWeightingKPCA)方法计算环境-中心同质度(HDHomogeneityDegree)：该方法抑制了CRF像元中的病态或异常数据，再对其进行主成分分析，提高主成分准确率；将nCRF各像元在中心主成分投影计算其与CRF中心同质度。其次基于该环境-中心多特征差异对nCRF抑制校正，构建一种同质度校正抑制模型：该模型在同质元素互抑制的同时，避免异质像元间相互作用及轮廓成份自抑制，实现微光图像中的噪声、纹理抑制，以及复杂自然场景下夜视图像的显著轮廓背景分离。The technical solution that realizes the object of the present invention is: at first research a kind of feature vector angle matching (FAMFeaturevectorAngleMatching) weighted KPCA (WKPCAWeightingKPCA) method calculation environment-center homogeneity degree (HDHomogeneityDegree) in high-dimensional feature space: this method has suppressed CRF pixel The pathological or abnormal data in the data, and then perform principal component analysis to improve the accuracy of the principal components; calculate the homogeneity between nCRF and the CRF center by projecting each pixel of nCRF on the central principal component. Secondly, based on the environment-center multi-feature difference to nCRF suppression correction, a homogeneity correction suppression model is constructed: this model avoids the interaction between heterogeneous pixels and the self-suppression of contour components while homogeneous elements suppress each other, and realizes Noise, texture suppression in low-light images, and remarkable silhouette-background separation in night-vision images of complex natural scenes.

本发明与现有技术相比，其显著优点：Compared with the prior art, the present invention has significant advantages:

（1）本发明针对微光图像多维特征分析，首次提出一种WKPCA算法，在高维特征空间基于特征向量角匹配准则对CRF协方差矩阵加权，有效削弱异常特征数据和噪声干扰，提高CRF主成分分析的准确性。(1) The present invention proposes a WKPCA algorithm for the first time for the multi-dimensional feature analysis of low-light images, which weights the CRF covariance matrix based on the feature vector angle matching criterion in the high-dimensional feature space, effectively weakens the abnormal feature data and noise interference, and improves the CRF main force. The accuracy of compositional analysis.

（2）本发明创新性的定义一种同质度概念及其计算方法，采用WKPCA算法进行CRF与nCRF区域同质性分析，在高维特征空间nCRF各像元特征向量投影到CRF中心主成分，计算两者同质度。该方法能够更全面准确的表针环境-中心的差异。(2) The invention innovatively defines a concept of homogeneity and its calculation method. The WKPCA algorithm is used to analyze the homogeneity of CRF and nCRF regions, and the eigenvectors of each pixel in nCRF are projected to the central principal component of CRF in the high-dimensional feature space. , to calculate the homogeneity of the two. This method enables a more comprehensive and accurate gauge needle environment-center difference.

（3）本发明提出一种WKPCA同质度校正nCRF抑制模型，根据nCRF各像元与CRF中心同质度对其抑制量校正。该模型更有效排除噪声干扰、削弱背景细节的同时，更精确的提高同质成份的抑制程度，降低异质像元、轮廓元素间的相互抑制，提高轮廓整体响应及减少轮廓间断的可能性。(3) The present invention proposes a WKPCA homogeneity corrected nCRF suppression model, and the suppression amount is corrected according to the homogeneity of each nCRF pixel and the CRF center. This model can more effectively eliminate noise interference and weaken background details, and at the same time, it can more accurately improve the degree of suppression of homogeneous components, reduce the mutual suppression between heterogeneous pixels and contour elements, improve the overall response of the contour and reduce the possibility of contour discontinuity.

附图说明Description of drawings

图1是nCRF距离衰减函数。Figure 1 is the nCRF distance decay function.

图2是WKPCA同质度校正nCRF抑制微光图像显著轮廓提取方法流程图。Fig. 2 is a flow chart of the method for extracting the salient contours of WKPCA homogeneity correction nCRF suppressed low-light images.

图3是多维特征空间非经典感受野结构。Figure 3 is a multi-dimensional feature space non-classical receptive field structure.

图4是以真实目标轮廓像元为中心的nCRF结构放大图，及nCRF中各像元特征向量与CRF中心均值向量FAM权值三维图。Figure 4 is an enlarged view of the nCRF structure centered on the real target contour pixel, and a three-dimensional diagram of the FAM weight of each pixel feature vector and CRF center mean vector in nCRF.

图5是PCA、KPCA、WKPCA算法对高斯数据的分类效果比较。Figure 5 is a comparison of the classification effects of PCA, KPCA, and WKPCA algorithms on Gaussian data.

图6是PCA、KPCA、WKPCA算法同质性分析结果比较。Figure 6 is a comparison of the homogeneity analysis results of PCA, KPCA, and WKPCA algorithms.

图7是背景、目标、轮廓中心的WKPCA同质度校正抑制分析。行向从上到下分别是原图、WKPCA同质度值输出、Gabor能量极值输出以及有效抑制输出（同质度值与Gabor能量极值乘积）。Figure 7 is a WKPCA homogeneity-corrected inhibition analysis of background, target, and contour centers. From top to bottom, the rows are the original image, WKPCA homogeneity value output, Gabor energy extreme value output and effective suppression output (the product of homogeneity value and Gabor energy extreme value).

图8是各模型对复杂自然场景下微光图像轮廓提取效果比较。Figure 8 is a comparison of the contour extraction effects of various models on low-light images in complex natural scenes.

具体实施方式detailed description

下面结合附图对本申请做进一步说明。The application will be further described below in conjunction with the accompanying drawings.

1、nCRF环境抑制：1. nCRF environmental suppression:

二维Gabor函数能有效地描述视皮层简单细胞的感受野剖面，通过奇偶对简单感受野滤波器的反应模（Gabor能量），能很好地模拟典型复杂细胞的基本特性。这些复杂细胞可以看成局部方位能量算子，用复杂细胞活动的最大值可以对图形边与线进行准确定位，因此本发明通过Gabor能量来模拟复杂细胞的响应。二维Gabor滤波器表示如下。The two-dimensional Gabor function can effectively describe the receptive field profile of simple cells in the visual cortex, and can well simulate the basic characteristics of typical complex cells through the response mode (Gabor energy) of parity to simple receptive field filters. These complex cells can be regarded as local azimuth energy operators, and the maximum value of complex cell activities can be used to accurately locate the edges and lines of graphics. Therefore, the present invention uses Gabor energy to simulate the response of complex cells. The two-dimensional Gabor filter is expressed as follows.

其中θ是CRF的偏好方位；是相差；由和分别表示奇偶Gabor滤波器；长宽比γ决定高斯椭圆的离心率；λ是波长；高斯标准差σ是决定了CRF作用面积；σ/λ表示空间频率带宽。in θ is the preferred orientation of the CRF; is the difference; Depend on and Represents the odd-even Gabor filter; the aspect ratio γ determines the eccentricity of the Gaussian ellipse; λ is the wavelength; the Gaussian standard deviation σ determines the CRF area; σ/λ represents the spatial frequency bandwidth.

简单细胞的响应为Gabor函数与输入图像I的卷积。由正交简单细胞响应定义CRF刺激的神经元响应E_σ(x,y;θ_i)。Simple Cell Responses is the convolution of the Gabor function with the input image I. The CRF-stimulated neuronal response E _σ (x,y; θ _i ) is defined by an orthogonal simple cellular response.

${E E.}_{σ σ}^{22} ((x x,, y the y;; {θ θ}_{i i})) = = {r r}_{λσ λσ 00}^{22} ((x x,, y the y;; {θ θ}_{22})) + + {r r}_{λσπ λσπ / / 22}^{22} ((x x,, y the y;; {θ θ}_{i i})) - - - - - - ((33))$

其中Gabor能量被分为N_θ个响应方位：θ_i=iπ/N_θ,i=0,1,…,N_θ-1。根据[9]设置参数：N_θ=12，γ=0.5，σ/λ=0.56，选择σ作为自变量。CRF的响应输出为各方位Gabor能量极值。The Gabor energy is divided into N _θ response orientations: θ _i =iπ/N _θ , i=0,1,…,N _θ -1. Set the parameters according to [9]: N _θ = 12, γ = 0.5, σ/λ = 0.56, and choose σ as the independent variable. The response output of CRF is the Gabor energy extremum of each orientation.

E_σ(x,y)=max{E_σ(x,y;θ_i)|i=0,1,…,N_θ-1}(4)E _σ (x,y)=max{E _σ (x,y;θ _i )|i=0,1,…,N _θ -1}(4)

生物研究表明V1区细胞的nCRF主要是抑制性的，这种环境抑制作用旨在分离同质区域，突显轮廓。在nCRF模型中采用距离衰减函数ω_σ(x,y)描述环境抑制的空间结构。ω_σ(x,y)定义为半波校正和L1规范的同轴高斯差函数DOG_σ,ρσ(x,y)。Biological studies have shown that the nCRF of cells in the V1 region is predominantly inhibitory, and that this environmental inhibition aims to isolate homogeneous regions and highlight contours. In the nCRF model, the distance decay function ω _σ (x, y) is used to describe the spatial structure of environmental suppression. ω _σ (x,y) is defined as the coaxial difference of Gaussian function DOG _σ,ρσ (x,y) with half-wave correction and L1 norm.

${DOG DOG}_{σ σ,, ρσ ρσ} ((x x,, y the y)) = = \frac{11}{\sqrt{22 π π {((ρσ ρσ))}^{22}}} exp exp ((- - \frac{{x x}^{22} + + {y the y}^{22}}{22 {((ρσ ρσ))}^{22}})) - - \frac{11}{\sqrt{22 π π {σ σ}^{22}}} exp exp ((- - \frac{{x x}^{22} + + {y the y}^{22}}{22 {σ σ}^{22}})) - - - - - - ((55))$

${| | z z | |}^{+ +} = = \{\begin{matrix} 00 & z z < < 00 \\ z z & z z &GreaterEqual; &Greater Equal; 00 \end{matrix} - - - - - - ((66))$

抑制形式t_σ(x,y)定义为环形背景中各像元的距离衰变局部方向能量。The suppression form t _σ (x,y) is defined as the distance-decayed local direction energy of each pixel in the ring background.

t_σ(x,y)=ω_σ(x,y)*E_σ(x,y)(7)t _σ (x,y)=ω _σ (x,y)*E _σ (x,y)(7)

如附图1所示，|DOG|⁺确保ω_σ(x,y)只作用于环境区域。σ和ρσ分别决定了CRF面积S_c与nCRF面积S_nc。生物实验表明nCRF直径能够达到CRF直径的2-5倍，因此取ρ=4。通过DOG_σ,kσ(x,y)=0，可以得到CRF半径r_c和nCRF半径ρr_c。As shown in Fig. 1, |DOG| ⁺ ensures that ω _σ (x,y) only acts on the environment region. σ and ρσ respectively determine the CRF area S _c and the nCRF area S _nc . Biological experiments show that the diameter of nCRF can reach 2-5 times the diameter of CRF, so ρ=4. By DOG _{σ, kσ} (x, y)=0, the CRF radius r _c and nCRF radius ρr _c can be obtained.

${r r}_{c c} = = \sqrt{22 ln ln ρ ρ / / {ρ ρ}^{22} - - 11} \cdot &Center Dot; ρσ ρσ - - - - - - ((88))$

nCRF环境抑制输出即为：The nCRF environmental suppression output is then:

R_σ(x,y)=|aE_σ(x,y)-bt_σ(x,y)|⁺ R _σ (x,y)=|aE _σ (x,y)-bt _σ (x,y)| ⁺

(9)(9)

2、同质度校正抑制2. Homogeneity correction suppression

如附图2所示，为提高环境-中心抑制的准确性，使得与中心同质的区域提供较大抑制量、异质的区域不对中心抑制，以及轮廓区域自抑制减少，本发明在nCRF模型基础上，对各环境像元与中心进行同质性分析，提出一种同质度计算方法，对环境抑制成份校正，实现强噪声干扰下微光图像背景纹理抑制和显著轮廓提取。As shown in Figure 2, in order to improve the accuracy of the environment-center suppression, so that the homogeneous region with the center provides a larger amount of suppression, the heterogeneous region does not suppress the center, and the self-inhibition of the contour region decreases, the present invention is based on the nCRF model Based on the homogeneity analysis of each environmental pixel and center, a homogeneity calculation method is proposed to correct the environmental suppression components, and realize background texture suppression and prominent contour extraction of low-light images under strong noise interference.

为计算环境各像元与中心同质度δ_w(x,y)，在多维特征空间，拟采用KPCA方法对CRF中心像元进行主成分分析，并将nCRF中各像元在CRF主成分投影计算两者差异性。但由于CRF中心存在病态或异常数据(噪声数据、背景中夹杂的目标数据、目标中夹杂的背景数据等)，导致主成分提取不准确，因此本发明提出一种特征向量角匹配(FAM)加权KPCA方法(WKPCA)，抑制病态或异常数据，提高主成分分析准确率。In order to calculate the homogeneity δ _w (x, y) between each pixel in the environment and the center, in the multi-dimensional feature space, it is proposed to use the KPCA method to conduct principal component analysis on the central pixel of the CRF, and project each pixel in the nCRF on the principal component of the CRF Calculate the difference between the two. However, due to the existence of pathological or abnormal data (noise data, target data mixed in the background, background data mixed in the target, etc.) in the center of the CRF, the principal component extraction is inaccurate, so the present invention proposes a feature vector angle matching (FAM) weighted The KPCA method (WKPCA) suppresses pathological or abnormal data and improves the accuracy of principal component analysis.

2.1多维特征分析2.1 Multidimensional feature analysis

上述环形nCRF结构在各方位上同等抑制，会导致异质像元及轮廓元素之间相互抑制，容易削弱轮廓响应及破坏轮廓完整性。蝶形抑制模型就是采用一种环境-中心方位差加权方法，在抑制纹理细节的同时减少共线轮廓元素间的相互作用，其对高质量可见光图像轮廓提取性能优于各向同性、各向异性模型，但是仍存在异向异质像元互抑制作用。而且相比于高质量可见光图像，微光图像中边缘受噪声干扰Gabor方向能量不准确甚至无明显方向性，因此只具有方向差加权的蝶形抑制无法很好提取微光图像轮廓。The above-mentioned circular nCRF structure is equally suppressed in all directions, which will lead to mutual suppression between heterogeneous pixels and contour elements, which will easily weaken the contour response and destroy the contour integrity. The butterfly suppression model uses an environment-center azimuth weighting method to suppress texture details while reducing the interaction between collinear contour elements. Its performance in contour extraction of high-quality visible light images is better than isotropic and anisotropic model, but there is still mutual inhibition of heterogeneous pixels. Moreover, compared with high-quality visible light images, the edges in low-light images are disturbed by noise, and the Gabor direction energy is inaccurate or even has no obvious directionality. Therefore, the butterfly suppression with only direction difference weighting cannot extract low-light image contours well.

生物实验表明当CRF和nCRF内的图形特征一致时抑制作用最强，而两者存在差异时抑制作用减弱或消失，这种相关特征包括方位、空间频率和对比度等。因此为弥补单一方向差加权抑制的不足，本发明采用多维特征分析，在特征空间更全面准确的计算nCRF与CRF差异。为有效表针微光图像特征，同时减少计算冗余，选取各像素5×5邻域内的4个旋转不变的统计特征以及方向作为像素的特征集，其中4个统计特征包括空间频率、对比度、灰度共生统计熵和灰度共生统计相关性。方向表针梯度信息；空间频率和对比度反映局部灰度统计分布；当像元间距设为1或2时，灰度共生统计特性能够描述局部高频信息，而且不同于Gabor滤波器采用高斯函数对局部区域加权，灰度共生统计特性假设邻域窗口内为均匀分布，因而对局部噪声不敏感。因此融合多种特征形成更具表针性的特征集。Biological experiments show that the inhibitory effect is the strongest when the graphic features in CRF and nCRF are consistent, and the inhibitory effect is weakened or disappears when there are differences between the two, such related features include orientation, spatial frequency and contrast. Therefore, in order to make up for the deficiency of single-directional difference weighted suppression, the present invention adopts multi-dimensional feature analysis to more comprehensively and accurately calculate the difference between nCRF and CRF in the feature space. In order to effectively feature low-light images of needles and reduce computational redundancy, four rotation-invariant statistical features and directions in the 5×5 neighborhood of each pixel are selected as the feature set of the pixel. The four statistical features include spatial frequency, contrast, Gray-level co-occurrence statistical entropy and gray-level co-occurrence statistical correlation. Direction pointer gradient information; spatial frequency and contrast reflect local gray level statistical distribution; when the pixel spacing is set to 1 or 2, gray level co-occurrence statistics can describe local high-frequency information, and it is different from Gabor filter using Gaussian function to local gray level Area weighting, gray-level co-occurrence statistical characteristics assume that the neighborhood window is uniformly distributed, so it is not sensitive to local noise. Therefore, multiple features are fused to form a more representative feature set.

定义fx＝[O_c,SF_c,CON_c,cooE_c,cooCOR_c]^T、fy＝[O_nc,SF_nc,CON_nc,cooE_nc,cooCOR_nc]^T分别表示CRF和nCRF内各像元特征向量。这里O_c、O_nc、SF_c、SF_nc、CON_c、CON_nc、cooE_c、cooE_nc、cooCOR_c和cooCOR_nc分别表示CRF和nCRF内各像元归一化的偏好方向、空间频率、对比度及灰度共生统计熵和相关性。Define fx＝[O _c , SF _c , CON _c , cooE _c , cooCOR _c ] ^T , fy＝[O _nc , SF _nc , CON _nc , cooE _nc , cooCOR _nc ] ^T represent the features of each pixel in CRF and nCRF respectively vector. Here O _c , On _nc , SF _c , SF _nc , CON _c , CON _nc , cooE _c , cooE _nc , cooCOR _c and cooCOR _nc represent the normalized preference direction, spatial frequency and contrast of each pixel in CRF and nCRF respectively And gray level co-occurrence statistical entropy and correlation.

2.2同质性分析2.2 Homogeneity analysis

为计算特征空间中nCRF各环境像元与CRF中心的差异，可以采用主成分分析(PCA)算法，即计算中心像元特征集的主成分，并将环境像元特征向量在主成分投影计算两者同质度。线性PCA算法要求数据服从高斯分布，该条件对真实场景很难满足；KPCA算法利用各维特征间的相关性，将线性空间特征信号映射到高维特征空间中进行差异检测，可以挖掘各特征间的非线性信息，更好的区分轮廓和背景。In order to calculate the difference between each environment pixel of nCRF and the CRF center in the feature space, the principal component analysis (PCA) algorithm can be used, that is, the principal component of the feature set of the center pixel is calculated, and the feature vector of the environment pixel is projected on the principal component to calculate two homogeneity. The linear PCA algorithm requires the data to obey the Gaussian distribution, which is difficult to satisfy in the real scene; the KPCA algorithm uses the correlation between the features of each dimension to map the linear space feature signal to the high-dimensional feature space for difference detection, and can mine the gap between the features. Non-linear information, better distinguish between contours and backgrounds.

KPCAKPCA

KPCA算法以检测点为中心，计算其CRF和nCRF数据高维特征空间差异。设子空间维度为P，定义P×N_c矩阵表示CRF多维特征数据，样本fx_i为P维特征向量fx＝[O_c,SF_c,CON_c,cooE_c,cooCOR_c]^T，N_c为CRF内像元数目；P×N_nc矩阵表示nCRF多维特征数据，样本fy_j为P维特征向量fy＝[O_nc,SF_nc,CON_nc,cooE_nc,cooCOR_nc]^T，N_nc为nCRF内像元数目。利用非线性映射函数φ将FX、FY映射到高维特征空间后，C_φ定义为CRF高维特征空间协方差：The KPCA algorithm takes the detection point as the center, and calculates the difference between the high-dimensional feature space of its CRF and nCRF data. Let the dimension of the subspace be P, and define the P×N _c matrix Represents CRF multi-dimensional feature data, sample fx _i is P-dimensional feature vector fx=[O _c ,SF _c ,CON _c ,cooE _c ,cooCOR _c ] ^T , N _c is the number of pixels in CRF; P×N _nc matrix Represents nCRF multi-dimensional feature data, sample fy _j is P-dimensional feature vector fy=[O _nc , SF _nc , CON _nc , cooE _nc , cooCOR _nc ] ^T , N _nc is the number of pixels in nCRF. After using the nonlinear mapping function φ to map FX and FY to the high-dimensional feature space, C _φ is defined as the covariance of the CRF high-dimensional feature space:

${C C}_{φ φ} = = \frac{11}{{N N}_{c c}} {Σ Σ}_{i i = = 11}^{{N N}_{c c}} ((φ φ (({fx fx}_{i i})) - - {μ μ}_{φ φ})) {((φ φ (({fx fx}_{i i})) - - {μ μ}_{φ φ}))}^{T T} - - - - - - ((1010))$

其中V_φ为C_φ非零特征值Λ_φ对应的特征向量。in V _φ is the eigenvector corresponding to the non-zero eigenvalue Λ _φ of C _φ .

nCRF各像元与CRF中心同质度δ(r)定义为The homogeneity δ(r) between each pixel of nCRF and the center of CRF is defined as

${δ δ}^{- - 11} ((r r)) = = {((φ φ ((r r)) - - {μ μ}_{φ φ}))}^{T T} {V V}_{φ φ} {V V}_{φ φ}^{T T} ((φ φ ((r r)) - - {μ μ}_{φ φ})) - - - - - - ((1111))$

其中r∈FY为nCRF各像元特征向量，即在高维特征空间计算环境像元特征在中心主成分的投影特性。Where r∈FY is the feature vector of each pixel in nCRF, that is, the projection characteristics of the environment pixel features in the central principal component are calculated in the high-dimensional feature space.

WKPCAWKPCA

KPCA能够更有效的提取和利用数据的非线性特征，但KPCA法也存在一定的缺陷，即若实际中心数据为病态分布(噪声干扰)、背景成份中目标点数目或目标成份中背景点数目较多时，高维特征空间协方差矩阵不能完全描述目标或背景数据。由式10可看出：CRF相关矩阵C_φ中各像元的权重相等，但若中心数据为病态分布、中心组成为背景数据目标点参杂或目标数据背景点参杂，C_φ就不能完全描述CRF数据分布，主成分分析不准确。所以高维特征空间检测点邻域的CRF协方差矩阵C_φ中各像素均需引入权重因子，抑制或降低目标数据中的背景点或异常点、背景数据中的目标点或异常点对C_φ的影响，C_φ就能更准确的描述目标、背景数据分布特性。KPCA can more effectively extract and utilize the nonlinear characteristics of data, but KPCA method also has certain defects, that is, if the actual central data is ill-conditioned distribution (noise interference), the number of target points in the background component or the number of background points in the target component is relatively small Many times, the high-dimensional feature space covariance matrix cannot fully describe the target or background data. It can be seen from Equation 10 that the weights of each pixel in the CRF correlation matrix C _φ are equal, but if the central data is ill-conditioned, the central composition is mixed with background data target points or target data background points are mixed, C _φ cannot be completely To describe the distribution of CRF data, principal component analysis is not accurate. Therefore, each pixel in the CRF covariance matrix C _φ of the detection point neighborhood in the high-dimensional feature space needs to introduce a weight factor to suppress or reduce the background points or abnormal points in the target data, and the target points or abnormal points in the background data _. C _φ can more accurately describe the distribution characteristics of target and background data.

因此本发明提出了一种WKPCA算法，该方法在高维特征空间以C_φ中各像元特征向量φ(fx_i)与数据中心向量均值μ_φ的特征角匹配(FAM)为准则，对C_φ中各像元特征向量引入对应的权重因子。即若像元特征向量与数据中心向量均值的夹角较小，表示两者特征信息相似，则该像元获得较大权值，反之亦然，如此去除或抑制C_φ中的异常数据。高维特征空间中加权CRF协方差矩阵C_φw表示为：Therefore the present invention proposes a kind of WKPCA algorithm, this method takes the feature angle matching (FAM) of each pixel feature vector φ(fx _i ) in C _φ and the data center vector mean value μ _φ as a criterion in the high-dimensional feature space, for C Each pixel feature vector in _φ introduces the corresponding weight factor. That is, if the angle between the pixel feature vector and the mean value of the data center vector is small, which means that the feature information of the two is similar, then the pixel gets a larger weight, and vice versa, so that the abnormal data in C _φ is removed or suppressed. The weighted CRF covariance matrix C _φw in the high-dimensional feature space is expressed as:

${C C}_{φw φw} = = \frac{11}{{N N}_{c c}} {Σ Σ}_{i i = = 11}^{{N N}_{c c}} {w w}_{{fx fx}_{i i}} ((φ φ (({fx fx}_{i i})) - - {μ μ}_{φ φ})) {((φ φ (({fx fx}_{i i})) - - {μ μ}_{φ φ}))}^{T T} / / {Σ Σ}_{i i = = 11}^{{N N}_{c c}} {w w}_{{fx fx}_{i i}} - - - - - - ((1212))$

式中为各像元的FAM权重因子：In the formula is the FAM weight factor of each pixel:

${w w}_{{fx fx}_{i i}} = = cos cos ((φ φ (({fx fx}_{i i})),, {μ μ}_{φ φ})) = = \frac{{φ φ}^{T T} (({fx fx}_{i i})) {μ μ}_{φ φ}}{| | | | {φ φ}^{T T} (({fx fx}_{i i})) | | | | | | | | {μ μ}_{φ φ} | | | |} = = \frac{\frac{11}{{N N}_{c c}} {Σ Σ}_{m m = = 11}^{{N N}_{c c}} k k (({fx fx}_{i i},, {fx fx}_{m m}))}{\sqrt{k k (({fx fx}_{i i},, {fx fx}_{i i}))} \sqrt{\frac{11}{{N N}_{c c}^{22}} {Σ Σ}_{m m = = 11}^{{N N}_{c c}} {Σ Σ}_{n no = = 11}^{{N N}_{c c}} k k (({fx fx}_{m m},, {fx fx}_{n no}))}} - - - - - - ((1313))$

i＝1,…,N_c，其中利用了核函数性质：k(fx,fy)＝<φ(fx),φ(fy)>。i=1,...,N _c , where the kernel function property is used: k(fx,fy)=<φ(fx),φ(fy)>.

WKPCA算法即计算高维特征空间中C_φw的特征向量V_φw及nCRF内各像元特征向量φ(r)在V_φw上的投影<V_φw,φ(r)>。The WKPCA algorithm calculates the feature vector V _φw of C _φw in the high-dimensional feature space and the projection <V _φw , φ(r)> of the feature vector φ(r) of each pixel in _nCRF on V φw.

定义高维特征空间加权数据 ${φ_{w} ({fx}_{i}) = \sqrt{w_{{fx}_{i}} / w} (φ ({fx}_{i}) - μ_{φ})}_{i = 1}^{N_{c}},$ 由式12可推出：definition High-dimensional feature space weighted data ${φ_{w} ({fx}_{i}) = \sqrt{w_{{fx}_{i}} / w} (φ ({fx}_{i}) - μ_{φ})}_{i = 1}^{N_{c}},$ It can be deduced from formula 12:

${C C}_{φw φw} = = \frac{11}{{N N}_{c c}} {Σ Σ}_{i i = = 11}^{{N N}_{c c}} {φ φ}_{w w} (({fx fx}_{i i})) {φ φ}_{w w} {(({fx fx}_{i i}))}^{T T} - - - - - - ((1414))$

为C_φw非零特征值构成的对角矩阵，为各特征值对应的特征向量。由式14可知，每个特征向量都在φ_w(FX)的度量空间内，因此可以表示成的线性组合： is a diagonal matrix composed of C _φw non-zero eigenvalues, is the eigenvector corresponding to each eigenvalue. It can be seen from formula 14 that each eigenvector are all in the metric space of φ _w (FX), so can be expressed as A linear combination of:

${V V}_{φw φw}^{l l} = = {Σ Σ}_{i i = = 11}^{{N N}_{c c}} {α α}_{i i}^{l l} {φ φ}_{w w} (({fx fx}_{i i})) = = {φ φ}_{w w} ((FX FX)) {α α}^{l l} - - - - - - ((1515))$

其中 $α^{l} = {[α_{1}^{l}, α_{2}^{l}, . . ., α_{N_{c}}^{l}]}^{T} .$ in $α^{l} = {[α_{1}^{l}, α_{2}^{l}, . . ., α_{N_{c}}^{l}]}^{T} .$

将14式和15式代入可得：Substitute formulas 14 and 15 into Available:

${N N}_{c c} {λ λ}_{w w}^{l l} {φ φ}_{w w} ((FX FX)) {α α}^{l l} = = {φ φ}_{w w} ((FX FX)) {φ φ}_{w w}^{T T} ((FX FX)) {φ φ}_{w w} ((FX FX)) {α α}^{l l} - - - - - - ((1616))$

将16式两边左乘且可得则核矩阵K_w的特征值为协方差矩阵C_φw的N_c倍，对应的特征向量利用核函数性质将K_w表示为：Multiply both sides of equation 16 to the left and Available Then the eigenvalue of the kernel matrix K _w is N _c times of the covariance matrix C _φw , and the corresponding eigenvector Using the properties of the kernel function, K _w is expressed as:

$\begin{matrix} {K K}_{w w} = = {D D.}_{w w} {N N}_{c c} {Λ Λ}_{φw φw} {D D.}_{w w}^{T T} = = {[[{φ φ}_{w w}^{T T} (({fx fx}_{i i})) {φ φ}_{w w} (({fx fx}_{j j}))]]}_{{N N}_{c c} \times \times {N N}_{c c}} & i i,, j j = = 11,, . . . . . .,, {N N}_{c c} \end{matrix}$

$= = \begin{matrix} {[[\frac{\sqrt{{w w}_{{fx fx}_{i i}} {w w}_{{fx fx}_{j j}}}}{w w} {((φ φ (({fx fx}_{i i})) - - {μ μ}_{φ φ}))}^{T T} ((φ φ (({fx fx}_{j j})) - - {μ μ}_{φ φ}))]]}_{{N N}_{c c} \times \times {N N}_{c c}} & i i,, j j = = 11,, . . . . . .,, {N N}_{c c} - - - - - - ((1717)) \end{matrix}$

$= = \begin{matrix} {[[\frac{\sqrt{{w w}_{{fx fx}_{i i}} {w w}_{{fx fx}_{j j}}}}{w w} ((k k (({fx fx}_{i i},, {fx fx}_{j j})) - - \frac{11}{{N N}_{c c}} {Σ Σ}_{m m = = 11}^{{N N}_{c c}} k k (({fx fx}_{i i},, {fx fx}_{m m})) - - \frac{11}{{N N}_{c c}} {Σ Σ}_{m m = = 11}^{{N N}_{c c}} k k (({fx fx}_{m m},, {fx fx}_{j j})) + + \frac{11}{{N N}_{c c}^{22}} {Σ Σ}_{m m = = 11}^{{N N}_{c c}} {Σ Σ}_{n no = = 11}^{{N N}_{c c}} k k (({fx fx}_{m m},, {fx fx}_{n no}))))]]}_{{N N}_{c c} \times \times {N N}_{c c}} & i i,, j j = = 11,, . . . . . .,, {N N}_{c c} \end{matrix}$

由15式可得，nCRF像元特征向量φ(r)在高维特征空间加权特征向量V_φw上的投影为：From Equation 15, the projection of the nCRF pixel feature vector φ(r) on the weighted feature vector V _φw of the high-dimensional feature space is:

$< < {V V}_{φw φw},, φ φ ((r r)) > > = = {D D.}_{w w}^{T T} {φ φ}_{w w}^{T T} ((FX FX)) φ φ ((r r)) = = {D D.}_{w w}^{T T} {k k}_{w w} ((Z Z,, r r)) - - - - - - ((1818))$

式中 $\begin{matrix} k_{w} (Z, r) = φ_{w}^{T} (FX) φ (r) = {[\sqrt{w_{{fx}_{i}} / w} (k ({fx}_{i}, r) - \frac{1}{N_{c}} Σ_{m = 1}^{N_{c}} k ({fx}_{m}, r))]}_{N_{c} \times 1} & i = 1, . . ., N_{c} \end{matrix}$ In the formula $\begin{matrix} k_{w} (Z, r) = φ_{w}^{T} (FX) φ (r) = {[\sqrt{w_{{fx}_{i}} / w} (k ({fx}_{i}, r) - \frac{1}{N_{c}} Σ_{m = 1}^{N_{c}} k ({fx}_{m}, r))]}_{N_{c} \times 1} & i = 1, . . ., N_{c} \end{matrix}$

高维特征空间nCRF各像元与中心CRF同质度δ_w(r)表示为：The homogeneity degree δ _w (r) between each pixel of the high-dimensional feature space nCRF and the central CRF is expressed as:

${δ δ}_{w w}^{- - 11} ((r r)) = = {((φ φ ((r r)) - - {μ μ}_{φ φ}))}^{T T} {V V}_{φw φw} {V V}_{φw φw}^{T T} ((φ φ ((r r)) - - {μ μ}_{φ φ}))$

$= = {((φ φ ((r r)) - - {μ μ}_{φ φ}))}^{T T} {φ φ}_{w w} ((FX FX)) {D D.}_{w w} {D D.}_{w w}^{T T} {φ φ}_{w w}^{T T} ((FX FX)) ((φ φ ((r r)) - - {μ μ}_{φ φ})) - - - - - - ((1919))$

$= = {(({k k}_{w w} ((Z Z,, r r)) - - \frac{11}{{N N}_{c c}} {Σ Σ}_{i i = = 11}^{{N N}_{c c}} {k k}_{w w} ((Z Z,, {fx fx}_{i i}))))}^{T T} {D D.}_{w w} {D D.}_{w w}^{T T} (({k k}_{w w} ((Z Z,, r r)) - - \frac{11}{{N N}_{c c}} {Σ Σ}_{i i = = 11}^{{N N}_{c c}} {k k}_{w w} ((Z Z,, {fx fx}_{i i}))))$

因此选择合适的核函数k构造正定的核矩阵K_w，计算其特征向量D_w，即可得出各环境像元与中心的同质度δ_w(x,y)。通过多次实验选用径向基核函数它能够表征特征向量间能量差异；同时采用FAM函数能够表征特征向量曲线的形状差异，如此WKPCA算法既考虑了像元间特征能量差异，也考虑了像元间特征曲线形状差异，更全面的区分多维特征数据。Therefore, choose an appropriate kernel function k to construct a positive definite kernel matrix K _w , and calculate its eigenvector D _w to obtain the homogeneity δ _w (x, y) between each environment pixel and the center. Radial basis kernel function was selected through multiple experiments It can represent the energy difference between feature vectors; at the same time, the FAM function can represent the shape difference of the feature vector curve, so that the WKPCA algorithm not only considers the difference in feature energy between pixels, but also considers the difference in the shape of feature curves between pixels, and can distinguish more comprehensively. Multidimensional feature data.

2.3同质度校正抑制2.3 Homogeneity correction suppression

仍然考虑抑制量随距离衰减特性，结合同质度δ_w(x,y)和距离衰减函数ω_σ(x,y)对nCRF中各像元抑制量校正，将nCRF模型中环境抑制T_σ(x,y)修订如下：Still considering the attenuation characteristics of the suppression amount with distance, combined with the homogeneity δ _w (x, y) and the distance attenuation function ω _σ (x, y) to correct the suppression amount of each pixel in the nCRF, the environmental suppression T _σ ( x,y) is revised as follows:

T_σ(x,y)=(δ_w(x,y)×ω_σ(x,y))*E_σ(x,y)T _σ (x,y)=(δ _w (x,y)×ω _σ (x,y))*E _σ (x,y)

(20)(20)

由于WKPCA更准确计算CRF中心主成分，目标与背景能够有效区分。对于检测点为背景内部像元情况，主要受与其同质区域的抑制，nCRF中有效抑制像元的数目和同质度值大，由于自然纹理Gabor方向能量极值大且分布杂乱，即使受距离衰减函数调制，CRF中心检测点仍获得较大抑制量而刺激响应被严重削弱；同样目标内部检测点主要受与其同质区域的抑制，有效抑制像元的数目取决于nCRF中同质区域大小(即目标大小)，目标区域同质度值大，其Gabor方向能量极值主要分布于目标内部和边缘，若目标较大则nCRF中同质区大，CRF中心目标像元获得较大抑制量而响应弱，若目标较小则nCRF中同质区小，CRF中心目标像元不受抑制，保留整体小目标；而对于显著目标轮廓检测点，由于WKPCA去除了异常干扰数据，CRF主成分与目标或背景特征吻合程度高，nCRF中目标内部的虚假边缘或背景纹理同质度值高，由于轮廓像元的多维特征区别于目标和背景、成像空间分辨率约束导致目标边缘产生混淆像元，其光谱特征与中心主成分产生偏差，轮廓像元本身同质度值较低，而其Gabor方向能量极值主要集中于边缘，因此中心轮廓像元主要受源自同一轮廓的像元抑制，有效抑制像元数目小且同质度值较小，加之距离衰减调制，轮廓像元几乎不受抑制而响应强。终上所述，WKPCA同质度校正抑制能够有效抑制环境纹理，获得较大轮廓响应，且最大程度减少轮廓自抑制，保留完整轮廓。Since WKPCA calculates the central principal component of CRF more accurately, the target and background can be effectively distinguished. For the case where the detection point is an internal pixel of the background, it is mainly suppressed by its homogeneous area. The number of effectively suppressed pixels and the homogeneity value in nCRF are large. Because the energy extreme value of the Gabor direction of the natural texture is large and the distribution is messy, even if it is affected by the distance When the attenuation function is modulated, the central detection point of the CRF still obtains a large amount of inhibition and the stimulus response is severely weakened; similarly, the internal detection point of the target is mainly inhibited by its homogeneous area, and the number of effectively inhibited pixels depends on the size of the homogeneous area in the nCRF ( That is, the size of the target), the homogeneity value of the target area is large, and its energy extremum in the Gabor direction is mainly distributed inside and at the edge of the target. If the response is weak, if the target is small, the homogeneous area in the nCRF is small, and the central target pixel of the CRF is not suppressed, and the overall small target is retained; and for the prominent target contour detection point, since the abnormal interference data is removed by WKPCA, the principal component of the CRF is consistent with the target Or the background features have a high degree of matching, the false edge inside the target in nCRF or the background texture homogeneity value is high, because the multi-dimensional features of the contour pixels are different from the target and the background, and the imaging spatial resolution constraints cause the target edge to produce confused pixels, its The spectral feature deviates from the central principal component, the homogeneity value of the contour pixel itself is low, and its energy extremum in the Gabor direction is mainly concentrated on the edge, so the central contour pixel is mainly suppressed by the pixels from the same contour, effectively suppressing The number of pixels is small and the homogeneity value is small, coupled with the distance attenuation modulation, the contour pixels are almost uninhibited and have a strong response. Finally, WKPCA homogeneity correction suppression can effectively suppress the environmental texture, obtain a larger contour response, and minimize the contour self-inhibition, retaining the complete contour.

下面结合实施例对本发明做进一步说明。The present invention will be further described below in conjunction with embodiment.

参照流程图2，本发明的具体实现步骤如下：With reference to flow chart 2, the specific implementation steps of the present invention are as follows:

步骤1，输入微光图像。Step 1, input the low-light image.

步骤2，设置非经典感受野模型参数：nCRF和CRF区域半径比ρ，高斯标准差σ，核函数尺度σ_k，调制系数a∈[0,1]、b∈[0,1]。根据式1及参数计算CRF和nCRF区域圆形半径r_c和ρr_c。Step 2. Set the non-classical receptive field model parameters: nCRF and CRF area radius ratio ρ, Gaussian standard deviation σ, kernel function scale σ _k , modulation coefficient a∈[0,1], b∈[0,1]. Calculate the circular radii r _c and ρr _c of the CRF and nCRF regions according to formula 1 and parameters.

${r r}_{c c} = = \sqrt{22 ln ln ρ ρ / / {ρ ρ}^{22} - - 11} \cdot &Center Dot; ρσ ρσ - - - - - - ((21 twenty one))$

步骤3，逐行扫描各个像素点，计算各像素点Gabor方向能量极值E_σ(x,y)；由式2计算nCRF区域内的距离加权函数ω_σ(x,y)。Step 3: Scan each pixel point by line, and calculate the Gabor direction energy extremum E _σ (x, y) of each pixel point; calculate the distance weighted function ω _σ (x, y) in the nCRF region by formula 2.

${DOG DOG}_{σ σ,, ρσ ρσ} ((x x,, y the y)) = = \frac{11}{\sqrt{22 π π {((ρσ ρσ))}^{22}}} exp exp ((- - \frac{{x x}^{22} + + {y the y}^{22}}{22 {((ρσ ρσ))}^{22}})) - - \frac{11}{\sqrt{22 π π {σ σ}^{22}}} exp exp ((- - \frac{{x x}^{22} + + {y the y}^{22}}{22 {σ σ}^{22}})) - - - - - - ((23 twenty three))$

步骤4，结合微光图像特性，计算图像各像素点特征向量，包括方向、空间频率、对比度、灰度共生统计熵和灰度共生统计相关性。Step 4: Combining the low-light image characteristics, calculate the feature vector of each pixel of the image, including direction, spatial frequency, contrast, gray-level co-occurrence statistical entropy and gray-level co-occurrence statistical correlation.

构建各像素点CRF区域内多维特征数据集其中fx＝[O_c,SF_c,CON_c,cooE_c,cooCOR_c]^T，N_c为CRF内像元数目，O_c、SF_c、CON_c、cooE_c和cooCOR_c为CRF区域内归一化的方向、空间频率、对比度、灰度共生统计熵和灰度共生统计相关性；nCRF区域内多维特征数据集其中fy＝[O_nc,SF_nc,CON_nc,cooE_nc,cooCOR_nc]^T，N_nc为nCRF内像元数目，O_nc、SF_nc、CON_nc、cooE_nc和cooCOR_nc为nCRF区域内归一化的方向、空间频率、对比度、灰度共生统计熵和灰度共生统计相关性。Construct a multi-dimensional feature data set in the CRF area of each pixel Where fx=[O _c , SF _c , CON _c , cooE _c , cooCOR _c ] ^T , N _c is the number of pixels in the CRF, O _c , SF _c , CON _c , cooE _c and cooCOR _c are the normalization in the CRF area Orientation, spatial frequency, contrast, gray-level co-occurrence statistical entropy and gray-level co-occurrence statistical correlation; multidimensional feature data set in nCRF area Where fy=[O _nc ,SF _nc ,CON _nc ,cooE _nc ,cooCOR _nc ] ^T , N _nc is the number of pixels in nCRF, On _nc , SF _nc , CON _nc , cooE _nc and cooCOR _nc are normalized in nCRF area Orientation, spatial frequency, contrast, gray level co-occurrence statistical entropy and gray level co-occurrence statistical correlation.

步骤5，基于WKPCA算法，采用径向基核函数，由式4计算CRF各像元的特征向量角匹配FAM值i＝1,…,N_c；根据FAM值对高维特征空间CRF各像元特征向量加权，计算加权后CRF多维特征数据集主成分，以及nCRF中各像元特征向量在CRF主成分上的投影，即由式6计算图像各像素点的nCRF区域内各像元与CRF中心同质度δ_w(r)。其中r∈FY，D_w为核矩阵K_w的特征向量。Step 5, based on the WKPCA algorithm, using the radial basis kernel function, calculate the eigenvector angle matching FAM value of each pixel of the CRF by formula 4 i=1,...,N _c ; according to the FAM value, weight the eigenvectors of each pixel in the high-dimensional feature space CRF, calculate the principal component of the weighted CRF multidimensional feature data set, and the eigenvectors of each pixel in nCRF on the CRF principal component Projection, that is, calculate the homogeneity δ _w (r) between each pixel and the CRF center in the nCRF area of each pixel in the image by formula 6. Among them r∈FY, D _w is the eigenvector of kernel matrix K _w .

${w w}_{{fx fx}_{i i}} = = \frac{\frac{11}{{N N}_{c c}} {Σ Σ}_{m m = = 11}^{{N N}_{c c}} k k (({fx fx}_{i i},, {fx fx}_{m m}))}{\sqrt{k k (({fx fx}_{i i},, {fx fx}_{i i}))} \sqrt{\frac{11}{{N N}_{c c}^{22}} {Σ Σ}_{m m = = 11}^{{N N}_{c c}} {Σ Σ}_{n no = = 11}^{{N N}_{c c}} k k (({fx fx}_{m m},, {fx fx}_{n no}))}} - - - - - - ((24 twenty four))$

$k k ((fx fx,, fy fy)) = = exp exp [[((- - {| | | | fx fx - - fy fy | | | |}^{22})) / / {22 σ σ}_{k k}^{22}]] - - - - - - ((2525))$

${δ δ}_{w w}^{- - 11} ((r r)) = = {(({k k}_{w w} ((Z Z,, r r)) - - \frac{11}{{N N}_{c c}} {Σ Σ}_{i i = = 11}^{{N N}_{c c}} {k k}_{w w} ((Z Z,, {fx fx}_{i i}))))}^{T T} {D D.}_{w w} {D D.}_{w w}^{T T} (({k k}_{w w} ((Z Z,, r r)) - - \frac{11}{{N N}_{c c}} {Σ Σ}_{i i = = 11}^{{N N}_{c c}} {k k}_{w w} ((Z Z,, {fx fx}_{i i})))) - - - - - - ((2626))$

${K K}_{w w} = = \begin{matrix} {[[\frac{\sqrt{{w w}_{{fx fx}_{i i}} {w w}_{{fx fx}_{j j}}}}{w w} ((k k (({fx fx}_{i i},, {fx fx}_{j j})) - - \frac{11}{{N N}_{c c}} {Σ Σ}_{m m = = 11}^{{N N}_{c c}} k k (({fx fx}_{i i},, {fx fx}_{m m})) - - \frac{11}{{N N}_{c c}} {Σ Σ}_{m m = = 11}^{{N N}_{c c}} k k (({fx fx}_{m m},, {fx fx}_{j j})) + + \frac{11}{{N N}_{c c}^{22}} {Σ Σ}_{m m = = 11}^{{N N}_{c c}} {Σ Σ}_{n no = = 11}^{{N N}_{c c}} k k (({fx fx}_{m m},, {fx fx}_{n no}))))]]}_{{N N}_{c c} \times \times {N N}_{c c}} & i i,, j j = = 11,, . . . . . .,, {N N}_{c c} - - - - - - ((2727)) \end{matrix}$

$\begin{matrix} {k k}_{w w} ((Z Z,, r r)) = = {φ φ}_{w w}^{T T} ((FX FX)) φ φ ((r r)) = = {[[\sqrt{{w w}_{{fx fx}_{i i}} / / w w} ((k k (({fx fx}_{i i},, r r)) - - \frac{11}{{N N}_{c c}} {Σ Σ}_{m m = = 11}^{{N N}_{c c}} k k (({fx fx}_{m m},, r r))))]]}_{{N N}_{c c} \times \times 11} & i i = = 11,, . . . . . .,, {N N}_{c c} - - - - - - ((2828)) \end{matrix}$

步骤6，结合同质度δ_w(x,y)和距离衰减函数ω_σ(x,y)对nCRF中各像元抑制量校正，由式9和式10计算图像各像素点的环境抑制量T_σ(x,y)及抑制输出R_σ(x,y)。其中δ_w(x,y)即为δ_w(r)。Step 6: Combining the homogeneity degree δ _w (x, y) and the distance attenuation function ω _σ (x, y) to correct the suppression amount of each pixel in the nCRF, and calculate the environmental suppression amount of each pixel in the image by Equation 9 and Equation 10 T _σ (x,y) and suppressed output R _σ (x,y). where δ _w (x, y) is δ _w (r).

T_σ(x,y)=(δ_w(x,y)×ω_σ(x,y))*E_σ(x,y)(29)T _σ (x,y)=(δ _w (x,y)×ω _σ (x,y))*E _σ (x,y)(29)

R_σ(x,y)=|aE_σ(x,y)-bT_σ(x,y)|⁺(30)R _σ (x,y)=|aE _σ (x,y)-bT _σ (x,y)| ⁺ (30)

步骤7，对输出R_σ(x,y)进行非极大值抑制，设置灰度阈值和轮廓长度阈值，对非极大值抑制结果进行二值化，得到最终的轮廓输出。Step 7: Perform non-maximum suppression on the output R _σ (x, y), set the gray threshold and contour length threshold, and binarize the non-maximum suppression results to obtain the final contour output.

本发明的效果可以通过以下仿真结果进一步说明：Effect of the present invention can be further illustrated by the following simulation results:

1、由图4可见，结构图中CRF窗口中主要为目标像元，其均值特征向量近似目标特性。因此FAM三维图中CRF窗口和nCRF环形窗口内目标像元的FAM值较大(白色)，背景像元的FAM值较小(黑色)。因此FAM权重因子能够有效削弱CRF协方差矩阵中的异常数据。实验中设置σ＝2.0，CRF窗口半径5，nCRF窗口半径20。1. It can be seen from Figure 4 that the CRF window in the structure diagram is mainly the target pixel, and its mean eigenvector approximates the target characteristics. Therefore, the FAM value of the target pixel in the CRF window and the nCRF ring window in the FAM 3D image is larger (white), and the FAM value of the background pixel is smaller (black). Therefore, the FAM weight factor can effectively weaken the abnormal data in the CRF covariance matrix. In the experiment, σ=2.0, CRF window radius 5, and nCRF window radius 20 are set.

2、由图5可见，对模拟数据仿真，测试比较了PCA、KPCA和WKPCA算法性能。相比于线性PCA，非线性KPCA一定程度的区分目标和背景，但是无法进一步对不同分布目标分类；而WKPCA不仅能够分离目标和背景，而且能够有效区分不同分布的目标。实验中设置σ_k＝0.5。2. It can be seen from Figure 5 that the performance of the PCA, KPCA and WKPCA algorithms was tested and compared for the simulated data simulation. Compared with linear PCA, nonlinear KPCA can distinguish targets and backgrounds to a certain extent, but cannot further classify targets of different distributions; while WKPCA can not only separate targets and backgrounds, but also effectively distinguish targets of different distributions. In the experiment, σ _k =0.5 is set.

3、如图6所示，WKPCA算法性能优势进一步表现在真实微光图像数据分析中。CRF中心以目标为主，KPCA比PCA主成分分析准确(a)。但是受异常数据(噪声、目标中的背景数据等)干扰，KPCA主成分产生偏差，导致目标、背景内部像元的同质度均较低；而目标轮廓像元多维统计特征包含目标和背景信息，其特征向量易与干扰主成分近似，导致目标轮廓的同质度较高(b)。WKPCA由于去除了异常数据的干扰，中心主成分计算准确，因而目标内部同质度较高；目标轮廓的灰度统计特征包含目标与背景信息，特征向量与中心主成分不一致而同质度稍低；背景同质度较低(c)。实验中设置σ＝2.0,σ_k＝0.5。3. As shown in Figure 6, the performance advantages of the WKPCA algorithm are further manifested in the analysis of real low-light image data. CRF centers on the target, and KPCA is more accurate than PCA principal component analysis (a). However, due to the interference of abnormal data (noise, background data in the target, etc.), the principal components of KPCA deviate, resulting in low homogeneity of the internal pixels of the target and the background; while the multidimensional statistical features of the target contour pixels contain target and background information , its eigenvectors are easy to approximate with the interference principal components, resulting in a higher homogeneity of the target profile (b). Because WKPCA removes the interference of abnormal data, the calculation of the central principal component is accurate, so the internal homogeneity of the target is relatively high; the gray-scale statistical features of the target outline include target and background information, and the eigenvector is inconsistent with the central principal component, so the homogeneity is slightly low ; lower background homogeneity (c). In the experiment, σ=2.0 and σ _k =0.5 are set.

采用式31计算PCA同质度，式中μ为CRF中心像元特征均值向量，v为采用PCA算法选定的主成分；采用式32计算KPCA同质度，w为KPCA算法选定的主成分；采用式19计算WKPCA同质度。PCA homogeneity is calculated by Equation 31, where μ is the feature mean vector of CRF central pixel, v is the principal component selected by PCA algorithm; KPCA homogeneity is calculated by Equation 32, and w is the principal component selected by KPCA algorithm ; Calculate WKPCA homogeneity using formula 19.

δ^-1(r)＝(r-μ)^Tvv^T(r-μ)(31)δ ^-1 (r) = (r-μ) ^T vv ^T (r-μ) (31)

${δ δ}^{- - 11} ((r r)) = = {((k k ((Z Z,, r r)) - - \frac{11}{{N N}_{c c}} {Σ Σ}_{i i = = 11}^{{N N}_{c c}} k k ((Z Z,, {fx fx}_{i i}))))}^{T T} {ww ww}^{T T} ((k k ((Z Z,, r r)) - - \frac{11}{{N N}_{c c}} {Σ Σ}_{i i = = 11}^{{N N}_{c c}} k k ((Z Z,, {fx fx}_{i i}))))$

(32)(32)

其中 $k (Z, r) = φ^{T} (FX) φ (r) = \begin{matrix} {[k ({fx}_{i}, r) - \frac{1}{N_{c}} Σ_{m = 1}^{N_{c}} k ({fx}_{m}, r)]}_{N_{c} \times 1} & i = 1, . . ., N_{c} . \end{matrix}$ in $k (Z, r) = φ^{T} (FX) φ (r) = \begin{matrix} {[k ({fx}_{i}, r) - \frac{1}{N_{c}} Σ_{m = 1}^{N_{c}} k ({fx}_{m}, r)]}_{N_{c} \times 1} & i = 1, . . ., N_{c} . \end{matrix}$

4、如图7所示，针对自然场景中微光图像验证WKPCA同质度校正抑制的合理有效性。对于背景纹理(列1-2)nCRF与CRF同质度高(行2)，Gabor方向能量极值杂乱(行3)，导致nCRF大部分像元对中心抑制，有效抑制像元数目大、各像元抑制值大(行4)，因此背景中心响应弱；对于目标内部(列3-4)nCRF中同一目标部分与CRF同质度高(行2)，Gabor方向能量极值集中于边缘与目标内部(行3)，导致nCRF中目标像元均以较大值对中心抑制，有效抑制像元数目取决于目标大小、抑制值大(行4)，因此目标中心响应弱；轮廓部分(列5-6)受WKPCA作用，nCRF中与CRF主成分一致的目标部分同质度高(行2)，轮廓部分同质度稍低，Gabor方向能量极值主要集中于边缘(行3)，目标内部的虚假边缘(列5行3)、背景纹理(列6行3)的Gabor能量较低，因此轮廓检测像元主要受源自同一轮廓的像元抑制，有效抑制像元数目小，且同质度值较小(行4)，加之距离校正，各像元抑制值小，因此轮廓中心响应强。实验中设置σ＝2.0,σ_k＝0.5。4. As shown in Figure 7, the reasonable effectiveness of WKPCA homogeneity correction suppression is verified for low-light images in natural scenes. For the background texture (column 1-2), the homogeneity between nCRF and CRF is high (row 2), and the energy extremum in the Gabor direction is disordered (row 3), which leads to the suppression of most pixels of nCRF to the center, effectively suppressing the large number of pixels and the The pixel suppression value is large (row 4), so the background center response is weak; for the same target part in the target interior (column 3-4) nCRF has high homogeneity with CRF (row 2), the Gabor direction energy extremum is concentrated on the edge and Inside the target (line 3), the target pixels in the nCRF are all suppressed to the center with a large value, and the number of effectively suppressed pixels depends on the size of the target, and the suppression value is large (line 4), so the response of the target center is weak; the contour part (column 5-6) Affected by WKPCA, the homogeneity of the target part consistent with the CRF principal component in nCRF is high (row 2), the homogeneity of the contour part is slightly lower, and the Gabor direction energy extremum is mainly concentrated on the edge (row 3), and the target The Gabor energy of the internal false edge (column 5, row 3) and background texture (column 6, row 3) is low, so the contour detection pixels are mainly suppressed by the pixels from the same contour, the number of effective suppression pixels is small, and the same The quality value is small (line 4), together with the distance correction, the suppression value of each pixel is small, so the response of the contour center is strong. In the experiment, σ=2.0 and σ _k =0.5 are set.

5、如图8所示，利用本申请方法提取微光图像中显著轮廓。预处理中采用低通滤波器降噪；后处理中进行非极大值抑制和二值化排除虚假边缘，以选择较高调制响应及较大长度的边缘作为轮廓输出。即使经过低通滤波仍存在噪声干扰，蝶形抑制残留了大量纹理和虚假边缘；线性PCA同质度校正抑制由于考虑了多特征，降低噪声干扰；KPCA同质度校正抑制采用非线性映射，提高了中心主成分准确度，更有效的区分目标和背景，进一步削弱了部分纹理和虚假边缘响应；WKPCA同质度校正抑制去除中心样本中的异常数据，最大程度的抑制同质区域响应，保留轮廓信息。实验中设置σ＝2.0,a＝1.0,b＝0.8,σ_k＝0.5。5. As shown in FIG. 8 , use the method of the present application to extract prominent contours in the low-light image. A low-pass filter is used to reduce noise in preprocessing; non-maximum suppression and binarization are used to eliminate false edges in postprocessing, so as to select edges with higher modulation response and longer length as contour output. Even after low-pass filtering, there is still noise interference, and butterfly suppression retains a large number of textures and false edges; linear PCA homogeneity correction suppression reduces noise interference due to consideration of multiple features; KPCA homogeneity correction suppression uses nonlinear mapping to improve Improve the accuracy of the central principal component, more effectively distinguish between the target and the background, and further weaken part of the texture and false edge response; WKPCA homogeneity correction suppresses the removal of abnormal data in the central sample, suppresses the homogeneous area response to the greatest extent, and preserves the contour information. In the experiment, σ=2.0, a=1.0, b=0.8, σ _k =0.5 are set.

Claims

1. the twilight image method for extracting remarkable configuration that a WKPCA homogeneity degree correction nCRF suppresses, it is characterised in that step is as follows:

Step 1, inputs twilight image；

Step 2, arranges non-classical receptive field model parameter: nCRF and CRF zone radius is than ρ, Gauss standard difference σ, kernel function yardstick σ_k, index of modulation a ∈ [0,1], b ∈ [0,1], calculate CRF and nCRF zoned circular radius r according to formula 1 and parameter_cWith ρ r_c；

r_{c} = \sqrt{2 l n ρ / ρ^{2} - 1} \cdot ρ σ - - - (1)

Step 3, progressively scans each pixel, calculates each pixel Gabor oriented energy extreme value E_σ(x,y)；The distance weighted function in nCRF region is calculated by formula 2(x, y) for pixel position in the picture；

ω_{σ_{1}} (x, y) = | {DOG}_{σ, ρ σ} (x, y) |^{+} / | | | {DOG}_{σ, ρ σ} (x, y) |^{+} | |_{1} - - - (2)

{DOG}_{σ, ρ σ} (x, y) = \frac{1}{\sqrt{2 π {(ρ σ)}^{2}}} \exp (- \frac{x^{2} + y^{2}}{2 {(ρ σ)}^{2}}) - \frac{1}{\sqrt{2 {πσ}^{2}}} \exp (- \frac{x^{2} + y^{2}}{2 σ^{2}}) - - - (3)

Step 4, calculates each pixel characteristic vector of twilight image, builds multi-dimensional feature data collection in each pixel CRF regionMulti-dimensional feature data collection in nCRF regionCharacteristic vector fx in CRF region_iElement is Gabor direction, spatial frequency, contrast, gray scale symbiosis statistical entropy and gray scale symbiosis statistic correlation, fx_i=[O_c,SF_c,CON_c,cooE_c,cooCOR_c]^T, i=1 ..., N_c, N_cFor pixel number in CRF, O_c、SF_c、CON_c、cooE_cAnd cooCOR_cFor Gabor direction normalized in CRF region, spatial frequency, contrast, gray scale symbiosis statistical entropy and gray scale symbiosis statistic correlation；Characteristic vector fy in nCRF region_j, j=1 ..., N_nc, fy_j=[O_nc,SF_nc,CON_nc,cooE_nc,cooCOR_nc]^T, N_ncFor pixel number in nCRF, O_nc、SF_nc、CON_nc、cooE_ncAnd cooCOR_ncFor Gabor direction normalized in nCRF region, spatial frequency, contrast, gray scale symbiosis statistical entropy and gray scale symbiosis statistic correlation；

Step 5, adopts WKPCA algorithm, calculates each pixel and CRF center homogeneity degree δ in the nCRF region of each pixel of image_w(r)；

Step 6, in conjunction with homogeneity degree δ_w(x, y) and range-attenuation functionEach pixel amount of suppression in nCRF is corrected, formula 9 and formula 10 calculates the environment amount of suppression T of each pixel of image_σ(x, y) and suppress output R_σ(x,y)；Wherein δ_w(x y) is δ_w(r)；

T_{σ} (x, y) = (δ_{w} (x, y) \times ω_{σ_{2}} (x, y)) * E_{σ} (x, y) - - - (9)

R_σ(x, y)=| aE_σ(x,y)-bT_σ(x,y)|⁺(10)

Step 7, to output R_σ(x, y) carries out non-maxima suppression, arranges gray threshold and profile length threshold value, non-maxima suppression result is carried out binaryzation, obtains final profile output.

2. the twilight image method for extracting remarkable configuration that WKPCA homogeneity degree according to claim 1 correction nCRF suppresses, it is characterized in that: the WKPCA algorithm in described step 5 is based on KPCA algorithm, in high-dimensional feature space CRF covariance matrix, introduced feature vectorial angle coupling FAM value is to each pixel characteristic vector weighting of CRF, and FAM value is the vector angle of each pixel characteristic vector of CRF and CRF characteristic data set vector average in high-dimensional feature space.

3. the twilight image method for extracting remarkable configuration that WKPCA homogeneity degree according to claim 1 and 2 correction nCRF suppresses, it is characterised in that: the step that the WKPCA algorithm in described step 5 calculates CRF characteristic data set main constituent is as follows:

First the characteristic vector angle coupling FAM value of each pixel of CRF is calculated by formula 4I=1 ..., N_c, and according to FAM value to each pixel characteristic vector weighting of high-dimensional feature space CRF；Calculate CRF multi-dimensional feature data collection main constituent after weighting again, and the projection in CRF main constituent of each pixel characteristic vector and δ in nCRF_w(r), r ∈ FY, D in formula 6_wFor nuclear matrix K_wCharacteristic vector, k () is Radial basis kernel function；

w_{{fx}_{i}} = \frac{\frac{1}{N_{c}} Σ_{m = 1}^{N_{c}} k ({fx}_{i}, {fx}_{m})}{\sqrt{k ({fx}_{i}, {fx}_{i})} \sqrt{\frac{1}{N_{c}^{2}} Σ_{m = 1}^{N_{c}} Σ_{n = 1}^{N_{c}} k ({fx}_{m}, {fx}_{n})}} - - - (4)

k (f x, f y) = \exp [(- | | f x - f y | |^{2}) / 2 σ_{k}^{2}] - - - (5)

δ_{w}^{- 1} (r) = {(k_{w} (Z, r) - \frac{1}{N_{c}} Σ_{i = 1}^{N_{c}} k_{w} (Z, {fx}_{i}))}^{T} D_{w} D_{w}^{T} (k_{w} (Z, r) - \frac{1}{N_{c}} Σ_{i = 1}^{N_{c}} k_{w} (Z, {fx}_{i})) - - - (6)

K_{w} = {[\frac{\sqrt{w_{{fx}_{i}} w_{{fx}_{j}}}}{w} (k ({fx}_{i}, {fx}_{j}) - \frac{1}{N_{c}} Σ_{m = 1}^{N_{c}} k ({fx}_{i}, {fx}_{m}) - \frac{1}{N_{c}} Σ_{m = 1}^{N_{c}} k ({fx}_{m}, {fx}_{j}) + \frac{1}{N_{c}^{2}} Σ_{m = 1}^{N_{c}} Σ_{n = 1}^{N_{c}} k ({fx}_{m}, {fx}_{n}))]}_{\begin{matrix} N_{c} \times N_{c} & i, j = 1, ..., N_{c} \end{matrix}} - - - (7)

k_{w} (Z, r) = φ_{w}^{T} (F X) φ (r) = {[\sqrt{w_{{fx}_{i}} / w} (k ({fx}_{i}, r) - \frac{1}{N_{c}} Σ_{m = 1}^{N_{c}} k ({fx}_{m}, r))]}_{\begin{matrix} N_{c} \times 1 & i = 1, ..., N_{c} \end{matrix}} - - - (8) .