WO2019201029A1 - Procédé et appareil de mise à jour de boîte candidate - Google Patents

Procédé et appareil de mise à jour de boîte candidate Download PDF

Info

Publication number
WO2019201029A1
WO2019201029A1 PCT/CN2019/077397 CN2019077397W WO2019201029A1 WO 2019201029 A1 WO2019201029 A1 WO 2019201029A1 CN 2019077397 W CN2019077397 W CN 2019077397W WO 2019201029 A1 WO2019201029 A1 WO 2019201029A1
Authority
WO
WIPO (PCT)
Prior art keywords
candidate
box
overlap
boxes
updated
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2019/077397
Other languages
English (en)
Chinese (zh)
Inventor
张锐
游骅
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huawei Technologies Co Ltd
Original Assignee
Huawei Technologies Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Huawei Technologies Co Ltd filed Critical Huawei Technologies Co Ltd
Publication of WO2019201029A1 publication Critical patent/WO2019201029A1/fr
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/285—Selection of pattern recognition techniques, e.g. of classifiers in a multi-classifier system
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition

Definitions

  • the present application relates to the field of computers, and in particular, to a method and apparatus for updating an alternative frame.
  • Non-Maximum Suppression is widely used in computer vision algorithms such as face recognition, edge detection or target detection.
  • the classifier is an important part of the NMS computer vision algorithm, which can be used to detect objects in an image, such as whether it is a face or the like. For an object in the image, the classifier generates a plurality of candidate boxes. The classifier calculates the classifier score for each candidate box. In order to accurately detect or identify an object, only one optimal candidate frame is reserved for one object, and the content of the optimal candidate frame is used as an object to be identified or detected.
  • the candidate box with the highest score of the classifier in all the candidate frames in the image is selected, and the overlapping area of the candidate frame with the highest score of the other candidate frame and the classifier is calculated by traversing, and the parts of the other candidate frames are deleted according to the overlapping area.
  • Optional box. The candidate box with the highest score of the classifier is again selected in the other candidate boxes that have not been deleted, and the traversal calculation of the overlap area is performed again until all the alternative boxes except the optimal candidate frame are deleted.
  • the number of candidate frames is very large, resulting in an increase in the amount of traversal calculation, which in turn increases the computational complexity of the NMS algorithm, reducing the efficiency of image recognition and detection.
  • the present application provides a method and apparatus for updating an alternative frame, which can improve the efficiency of updating an optional frame, thereby improving the efficiency of image recognition and detection.
  • the present application provides an alternative frame update method, including: obtaining, according to location information and a location vector of a first candidate frame, an overlap between a first candidate frame and a plurality of second candidate frames, An alternative box is an alternative box with the highest score of the classifier in the set of candidate boxes, and the second candidate box is an alternative box in the set of candidate boxes except the first candidate box, the position vector includes a plurality of second Position information of the candidate box; comparing the overlap degree and the overlap degree threshold of the first candidate box with the plurality of second candidate boxes to obtain an updated set of candidate boxes.
  • the degree of overlap between the first candidate frame and the plurality of second candidate frames can be obtained by one calculation.
  • the calculated overlap and overlap degree thresholds of the first candidate box and the second candidate box may be used to obtain an candidate box in the updated candidate box set.
  • the position vector can greatly reduce the time required to update the candidate box in the candidate box set, and the efficiency of updating the candidate box is improved, thereby improving the use of the candidate box in the present application.
  • the efficiency of the method's NMS method improves the efficiency of image recognition and detection.
  • the first candidate box is compared with the overlap degree and the overlap threshold of the plurality of second candidate boxes to obtain an updated set of candidate boxes, including: When the overlap degree of the second candidate box and the first candidate box in the second candidate box is less than or equal to the overlap degree threshold, determining that the second candidate box remains in the updated candidate box set; When the degree of overlap of the second candidate box and the first candidate box in the candidate box is greater than the overlap threshold, it is determined that the second candidate box is not retained in the updated set of candidate boxes.
  • Maintaining a second candidate frame with a degree of overlap with the first candidate frame that is less than or equal to the overlap degree threshold as an alternative frame in the updated candidate set of frames, and deleting the overlap with the first candidate frame is greater than the overlap threshold The second alternative box.
  • the candidate frame updating method further includes: using the location information of the first candidate frame in the updated candidate frame set and the updated location vector to obtain the updated candidate.
  • each target recognition frame can identify one target object.
  • the degree of overlap between the first candidate frame and the second candidate frame is obtained by using the position vector for each update of the candidate frame set, thereby reducing the time taken to identify the plurality of target objects in the image, and improving the plurality of recognition images. The productivity of the target object.
  • the set of candidate boxes includes a plurality of candidate boxes for identifying faces in the image.
  • the present application provides an alternative frame updating apparatus, including: an overlap acquiring module, configured to obtain a first candidate frame and multiple second devices according to location information and a location vector of the first candidate frame.
  • the overlap of the marquee, the first candidate box is the candidate box with the highest score of the classifier in the set of candidate boxes, and the second candidate box is the candidate box of the set of candidate boxes except the first candidate box.
  • the location vector includes location information of the plurality of second candidate frames, and an update module, configured to compare the overlap degree and the overlap threshold of the first candidate frame with the plurality of second candidate frames to obtain an updated candidate frame set.
  • the updating module is specifically configured to determine, when the overlap degree of the second candidate box and the first candidate box in the multiple second candidate frames is less than or equal to the overlap degree threshold
  • the second candidate box is reserved in the updated candidate box set; when the overlap degree of the second candidate box and the first candidate box in the plurality of second candidate boxes is greater than the overlap degree threshold, determining the second standby The marquee does not remain in the updated set of alternate boxes.
  • the overlap acquiring module is further configured to obtain the updated candidate frame by using the location information of the first candidate frame in the updated candidate set of frames and the updated location vector.
  • the update module is further configured to compare the overlap between the first candidate box in the updated set of candidate boxes and the overlap and overlap thresholds of the plurality of second candidate boxes in the updated set of candidate boxes To obtain the updated candidate box set again, until the overlap degree of the first candidate box in the updated candidate box set obtained by the overlap degree acquiring module and the second candidate box in the updated candidate box set All of which are greater than the overlap threshold, the update module is configured to stop updating the set of candidate boxes, and to use the first candidate box in the set of candidate boxes and the first candidate box in the set of candidate boxes after each update. , as the target recognition box.
  • the set of candidate boxes includes a plurality of candidate boxes for identifying faces in the image.
  • the second aspect and the candidate frame updating device in the implementation of the second aspect can also achieve the same technical effects as the candidate frame updating method in the above technical solution.
  • the present application provides an alternative frame updating apparatus, including a memory, a processor, and a program stored on the memory and executable on the processor, and the processor implements an alternative as in the foregoing technical solution when executing the program.
  • the frame update method can achieve the same technical effect as the candidate frame update method in the above technical solution.
  • the present application provides a storage medium on which a program is stored, and when the program is executed by the processor, an optional frame update method in the foregoing technical solution is implemented, and an optional block update in the foregoing technical solution can be achieved.
  • the same technical effect of the method is not limited to:
  • FIG. 3 is a schematic diagram showing changes of an alternative box for two screenings in NMS calculation according to an embodiment of the present application
  • FIG. 4 is a schematic diagram of an alternative frame in the image to be tested in the embodiment of the present application.
  • FIG. 5 is a schematic structural diagram of an apparatus for updating an alternative frame according to an embodiment of the present application.
  • FIG. 6 is a schematic structural diagram of hardware of an alternative frame updating device in an embodiment of the present application.
  • the present application provides an alternative frame updating method and apparatus, which can be applied to a scene of visual image processing.
  • a target recognition frame can be generated.
  • the content in the target recognition frame is the target object in the image under test.
  • a plurality of identification frames may be generated for one target object, and the plurality of identification frames are candidates for the target recognition frame, hereinafter referred to as an optional frame (including the first candidate frame and the first Two alternative boxes).
  • An optimal candidate box ie, the first candidate box
  • the target image may include multiple target objects, and multiple rounds of candidate boxes need to be filtered to obtain an optimal candidate box corresponding to all target objects.
  • the screening of each round of alternative boxes, except for the screening of the last round of alternate boxes, will be updated to get the set of candidate boxes that are filtered in the next round of candidate boxes.
  • the update of the candidate boxes in the set of candidate boxes may employ the alternate box update method in this application.
  • the candidate frame update in the embodiment of the present application may be implemented by adding a vector processing unit to the image processing device or adding a vector processing function to the existing unit in the image processing device.
  • the candidate box update in the embodiment of the present application is specifically applicable to a method of Non-Maximum Suppression (NMS).
  • NMS Non-Maximum Suppression
  • the NMS method can be used in target object recognition for visual image processing.
  • the image processing device may be specifically a non-maximum suppression device.
  • the non-maximum suppression device may collect non-maximum suppression data including the classifier score of the candidate frame in the measured image and the position information of the candidate frame.
  • the collected non-maximum suppression data can be transferred from the system double data rate (English: Double Data Rate, DDR) to the local storage space (ie, Local Memory) through integrated direct memory access (IDMA). )in.
  • the size of the local storage space is limited, and the non-maximum suppression data in the system DDR needs to be transferred to the local storage space in batches.
  • the unit implementing the vector processing function in the non-maximum suppression device may generate a position vector according to the vectorization instruction and the non-maximum value suppression data, and perform calculation of the position vector and other data until all the recognition target objects in the measured image are obtained. Or track the target recognition frame of the target object.
  • FIG. 1 is a flowchart of an alternate frame update method in an embodiment of the present application.
  • the candidate frame updating method may include step S101 and step S102.
  • step S101 the degree of overlap between the first candidate frame and the plurality of second candidate frames is obtained according to the location information of the first candidate frame and the location vector.
  • the first candidate box is an candidate box with the highest score of the classifier in the set of candidate boxes.
  • the second alternative box is an alternative box in the set of candidate boxes other than the first candidate box.
  • the set of candidate boxes includes a plurality of candidate boxes.
  • the location information can characterize the location of the candidate frame in the image being measured. For example, if the candidate frame is a rectangular frame, the position information may be the coordinates of the four vertices of the rectangular frame in the image to be measured, or the position information may be a pair of vertices on the diagonal line of the rectangular frame in the image to be tested. The coordinates in . Based on the location information of the two candidate boxes, the degree of overlap of the two candidate boxes can be calculated.
  • the location vector includes location information of a plurality of second candidate frames. That is to say, the position vector includes a plurality of elements, and one element is position information of a second alternative frame. Using the position information of the first candidate frame and the position vector, the degree of overlap of the first candidate frame and the plurality of second candidate frames can be calculated at one time.
  • step S102 the first candidate box is compared with the overlap degree and the overlap degree threshold of the plurality of second candidate boxes to obtain an updated set of candidate boxes.
  • the degree of overlap may be an Intersection Over Union (IOU) or a ratio of the area of the overlap to the area of the candidate frame, which is not limited herein.
  • the overlap degree threshold may be a overlap area threshold, or may be a ratio threshold of the overlap area to the candidate frame area, and is not limited herein.
  • the second candidate box with the degree of overlap with the first candidate box being less than or equal to the overlap degree threshold is reserved as an alternative box in the updated set of candidate boxes.
  • the second candidate box having a degree of overlap with the first candidate box that is greater than the overlap degree threshold is not retained in the updated set of candidate boxes.
  • the second candidate box with the overlap degree of the first candidate frame greater than the overlap degree threshold is deleted. That is, the second candidate box with the degree of overlap with the first candidate box being greater than the overlap threshold does not participate in the next round of candidate box screening.
  • the calculation amount required for updating the next round of the candidate box set is reduced, and the resources occupied by the calculation of the next round of the candidate box set are also reduced, and the efficiency of updating the candidate box set is improved, thereby further improving The efficiency of the NMS method.
  • the NMS method can be implemented by using the candidate frame update method of the embodiment of the present application. Specifically, the overlap degree calculation in the step S101 and the step S102 and the updated content of the candidate frame set can be repeated by using the updated candidate frame set until the update is performed. The degree of overlap of the first candidate block in the subsequent set of candidate blocks with the second candidate block in the updated set of candidate blocks is greater than the overlap threshold, and the updating of the set of candidate blocks may be stopped. All of the first candidate boxes obtained during the process of updating the candidate box multiple times can be used as the target recognition box.
  • the degree of overlap between the first candidate frame and the plurality of second candidate frames can be obtained by one calculation.
  • the calculated overlap and overlap degree thresholds of the first candidate box and the second candidate box may be used to obtain an candidate box in the updated candidate box set.
  • the location vector can greatly reduce the time required for updating the candidate box in the candidate box set, and the efficiency of updating the candidate box is improved, thereby improving the utilization of the embodiment of the present application.
  • the efficiency of the NMS method of the marquee update method improves the efficiency of image recognition and detection.
  • the following is an example of identifying a target object in a measured image by using a non-maximum suppression method, and explaining the details of the non-maximum suppression method using the candidate frame update method and the candidate frame update method therein.
  • FIG. 2 is a flowchart of a method for non-maximum value suppression in the embodiment of the present application. As shown in FIG. 2, the method of non-maximum value suppression may include steps S201 to S205.
  • step S201 in the candidate box set, the candidate box with the highest score of the classifier is selected as the first candidate box.
  • the initial set of candidate boxes includes all candidate boxes on the image to be tested.
  • the classifier calculates the classifier score for each candidate box. By inputting the measured image and the candidate box on the measured image into the classifier, the classifier score of each candidate frame on the measured image can be obtained.
  • the classifier can gradually improve the classification and scoring standards through machine learning.
  • the classifier score reflects the degree to which each candidate box matches the target object in the image being measured. The higher the classifier score of the candidate box, the higher the degree to which the candidate box matches the target object in the image under test.
  • Multiple candidate boxes may be generated for each target object in the image being measured.
  • multiple screenings are required.
  • a first candidate box is obtained in each screening, and the first candidate box is an alternative box that best matches the target object.
  • the candidate box with the highest score of the classifier can be selected as the first candidate box.
  • the candidate frame updating method and the NMS method in the embodiments of the present application can be applied to scenes such as face recognition, edge detection, or target detection.
  • scenes such as face recognition, edge detection, or target detection.
  • the set of candidate boxes contains a plurality of candidate boxes for identifying faces in the image.
  • the face in the image can be recognized more efficiently by using the candidate frame updating method and the NMS method in the embodiment of the present application.
  • step S202 a position vector is generated using the position information of the second candidate frame.
  • the second candidate box includes an alternative box in the set of candidate boxes except the first candidate box.
  • the 6 candidate boxes are D1, D2, D3, D4, D5 and D6.
  • the classifier score of the candidate box D1 is 0.98
  • the classifier score of the candidate box D2 is 0.81
  • the classifier score of the candidate box D3 is 0.69
  • the classifier score of the candidate box D4 is 0.48
  • the classifier score is 0.24 and the classifier score for candidate box D6 is 0.58.
  • the first candidate box is the candidate box D1
  • the second candidate box is the candidate box D2, the candidate box D3, the candidate box D4, the candidate box D5, and the candidate box D6.
  • Each location vector includes location information for a plurality of second candidate frames.
  • the position information of each second candidate box can be used as an element in the position vector.
  • the number of elements in the position vector can be preset, for example, according to a specific work scene and work requirements, and is not limited herein.
  • the number of second candidate boxes may be divisible by the number of elements preset by the position vector, ie, at least one position vector is generated. Then, in the candidate frame update process, the degree of overlap between the first candidate frame and all the second candidate frames is obtained according to the location information of the first candidate frame and the at least one location vector. For example, in the process of an alternate box update, the number of elements of the preset position vector is four, and the number of the second candidate frame is eight, and two position vectors can be generated.
  • the number of second candidate frames may not be divisible by the number of elements preset by the position vector, and based on the generation of the at least one position vector, there is also an integer single second candidate that does not form a position vector.
  • the degree of overlap of the first candidate frame with the second candidate frame involved in the location vector may be obtained according to the location information of the first candidate frame and the at least one location vector.
  • the location information of the first candidate frame and the location information of the integer single second candidate frame may be used to obtain the degree of overlap between the first candidate frame and the integer single second candidate frame. For example, in one screening, there are 8 second candidate boxes, and the number of elements of the preset position vector is 3, and position information of 2 position vectors and 2 single second candidate boxes is generated.
  • the second candidate frame may be sorted in descending order of the classifier score, and the number of elements per preset position vector may be second selected in descending order of the classifier score.
  • the position information of the box is used as a position vector.
  • the remainder obtained by quoting the number of the second candidate frames and the number of elements of the preset position vector is selected according to the order of the classifier scores of the second candidate frame from low to high, and the remaining second candidate frames are selected.
  • the position information of the remaining second candidate boxes is taken as the position information of the integer single second candidate frame. That is to say, if at least one position vector is generated, there is also position information of P second candidate frames in which the position vector is not formed, and P is a positive integer. Then, the location information of the P second candidate frames is the location information of the P second candidate frames with the lowest score of the classifier.
  • step S203 vector calculation is performed on the position information and the position vector of the first candidate frame to obtain the degree of overlap between the first candidate frame and the second candidate frame.
  • a vectorization instruction can be generated in the case where vector calculation is required, and vector calculation is triggered according to the vectorization instruction.
  • the vector calculation may calculate the degree of overlap of the first candidate box with the plurality of second candidate boxes in a position vector at one time. For example, let the number of second candidate boxes be 8, and the number of elements in the position vector be four. If the NMS method in the prior art is used, it is necessary to perform 8 calculations in order to obtain the overlap of the first candidate box and all the 8 second candidate boxes in one screening. With the NMS method in the embodiment of the present application, the degree of overlap between the first candidate box and all the eight second candidate boxes is obtained in one screening, and only two calculations are needed. The amount of calculation in the NMS method is greatly reduced, and the calculation time spent by the NMS method is reduced.
  • step S204 if the degree of overlap between the first candidate frame and the second candidate frame is not all greater than the overlap threshold, the candidate set of frames is updated to include the overlap with the first candidate frame being less than or A second alternative box equal to the overlap threshold.
  • step S205 if the degree of overlap between the first candidate box and the second candidate box is greater than the overlap degree threshold, the updating of the candidate box set is stopped.
  • multiple screenings may be required.
  • the first candidate box is selected in each screening, and the degree of overlap between the first candidate box and the second candidate box in the current screening is calculated.
  • the screening is stopped until the overlap between the first candidate box and all the second candidate boxes in the current screening is greater than the overlap threshold.
  • the first candidate box is obtained in each screening, that is, an alternative frame that best matches each of the target objects in the image to be measured is obtained.
  • a second candidate box with a degree of overlap with the first candidate box that is greater than the overlap threshold may be deleted in each screening. That is to say, the second candidate box whose overlap degree with the first candidate box in the current screening is greater than the overlap degree threshold is no longer involved in the calculation of the screening after the current screening. The calculation amount and the time taken for calculation are reduced, and the resources occupied by the calculation are also reduced, and the efficiency of the NMS method is further improved.
  • an index number (ie, Index) can be set for each candidate box, and an attribute (ie, Flags) is set for each candidate box, and the attribute of the candidate box indicates the candidate box and the first candidate in the current filter. Whether the overlap of the boxes is greater than the overlap threshold.
  • FIG. 3 is a schematic diagram showing changes of an alternative box for two screenings in NMS calculation in the embodiment of the present application. As shown in FIG. 3, an attribute of 0 indicates that the degree of overlap with the first candidate box is less than or equal to the overlap degree threshold, and an attribute of 1 indicates that the degree of overlap with the first candidate box is greater than the overlap threshold.
  • the attributes corresponding to the index numbers 0, 1, 2, and 3 are 0, 1, 1, and 0, respectively.
  • the candidate box with the attribute 1 is deleted, and the candidate box with the index numbers 0 and 3 is left in the mth filter.
  • FIG. 4 is a schematic diagram of an alternative frame in the image to be tested in the embodiment of the present application.
  • the test image includes a total of seven candidate boxes, which are an optional box D1, an optional box D2, an optional box D3, an optional box D4, an optional box D5, an optional box D6, and a standby device. Box D7.
  • the classifier scores of the candidate block D1, the candidate box D2, the candidate box D3, the candidate box D4, the candidate box D5, the candidate box D6, and the candidate box D7 are 0.98, 0.81, 0.69, 0.48, 0.58, respectively. 0.24 and 0.36.
  • the overlap threshold is the percentage of the overlap area to the area of the candidate frame, and the overlap threshold is 20%.
  • the seven candidate boxes can be arranged according to the classifier score, and the order of arrangement is the candidate box D1, the candidate box D2, the candidate box D3, the candidate box D6, the candidate box D4, the candidate box D7, and the candidate box. D5.
  • the candidate box D1 with the highest score of the classifier is selected as the first candidate box, the candidate box D2, the candidate box D3, the candidate box D6, the candidate box D4, the candidate box D7, and the preparation Box D5 is the second alternative box.
  • the first position vector includes position information of the candidate box D2, position information of the candidate box D3, and position information of the candidate box D6.
  • the second position vector includes position information of the candidate box D4, position information of the candidate box D7, and position information of the candidate box D5. Then, the calculation needs to be performed twice, and the first calculation results the overlap degree between the first candidate box D1 and the candidate box D2, the candidate box D3 and the candidate box D6 in the first position vector, and the second calculation is obtained. The degree of overlap of each of the first candidate box D1 and the candidate box D4, the candidate box D5, and the candidate box D7 in the second position vector.
  • the degree of overlap of the first candidate box D1 and the candidate box D2 is greater than 20%; the degree of overlap of the first candidate box D1 with the candidate box D3 is greater than 20%; the degree of overlap of the first candidate box D1 with the candidate box D6 Less than 20%; the degree of overlap of the first candidate box D1 and the candidate box D4 is greater than 20%; the degree of overlap of the first candidate box D1 with the candidate box D7 is less than 20%; the first candidate box D5 and the candidate box The overlap of D2 is less than 20%. Then, the candidate box D2 and the candidate box D3 are deleted.
  • the set of candidate boxes includes an alternative box D6, an alternate box D4, an alternate box D7, and an alternate box D5.
  • the candidate box D6 with the highest score of the classifier is selected as the first candidate box.
  • a position vector is generated, which includes the position information of the candidate box D4, the position information of the candidate box D7, and the position information of the candidate box D5.
  • the degree of overlap The degree of overlap of the first candidate block D6 and the candidate box D4 is less than 20%; the degree of overlap of the first candidate box D6 with the candidate box D7 is greater than 20%; the degree of overlap of the first candidate box D6 with the candidate box D5 More than 20%.
  • the candidate box D4 is deleted.
  • the candidate box D7 is selected as the first candidate box, and the degree of overlap between the first candidate box D7 and the candidate box D5 is calculated.
  • the degree of overlap of the first candidate block D7 with the candidate box D5 is less than 20%.
  • the candidate box D5 is also used as the first candidate box, and the update of the candidate box in the candidate box set is stopped, and the NMS calculation is completed.
  • the position information of one calculation vector and two single second candidate frames is generated in the first screening.
  • the seven candidate boxes can be arranged according to the classifier score, and the order of arrangement is the candidate box D1, the candidate box D2, the candidate box D3, the candidate box D6, the candidate box D4, the candidate box D7, and the candidate box. D5.
  • the candidate box D1 with the highest score of the classifier is selected as the first candidate box.
  • the location vector includes location information of the candidate frame D2, location information of the candidate frame D3, location information of the candidate frame D6, and location information of the candidate frame D4.
  • the location information of the two single second candidate frames is the location information of the candidate frame D7 and the location information of the candidate frame D5, respectively.
  • the first calculation calculates the degree of overlap between the first candidate box D1 and the candidate box D2, the candidate box D3, the candidate box D6 and the candidate box D4 in the position vector, and the second calculation results the first candidate box.
  • the degree of overlap of D1 with the candidate box D5 is calculated for the third time to obtain the degree of overlap between the first candidate box D1 and the candidate box D7.
  • the degree of overlap of the first candidate box D1 and the candidate box D2 is greater than 20%; the degree of overlap of the first candidate box D1 with the candidate box D3 is greater than 20%; the degree of overlap of the first candidate box D1 with the candidate box D6 Less than 20%; the degree of overlap of the first candidate box D1 and the candidate box D4 is greater than 20%; the degree of overlap of the first candidate box D1 with the candidate box D7 is less than 20%; the first candidate box D5 and the candidate box The overlap of D2 is less than 20%. Then, the candidate box D2 and the candidate box D3 are deleted.
  • the set of candidate boxes includes an alternative box D6, an alternate box D4, an alternate box D7, and an alternate box D5.
  • the candidate box D6 with the highest score of the classifier is selected as the first candidate box.
  • the degree of overlap of the first candidate box D6 with the candidate box D4, the candidate box D7, and the candidate box D5, respectively, is calculated.
  • the degree of overlap of the first candidate block D6 and the candidate box D4 is less than 20%; the degree of overlap of the first candidate box D6 with the candidate box D7 is greater than 20%; the degree of overlap of the first candidate box D6 with the candidate box D5 More than 20%.
  • the candidate box D4 is deleted.
  • the candidate box D7 is selected as the first candidate box, and the degree of overlap between the first candidate box D7 and the candidate box D5 is calculated.
  • the degree of overlap of the first candidate block D7 with the candidate box D5 is less than 20%.
  • the candidate box D5 is also used as the first candidate box, and the update of the candidate box in the candidate box set is stopped, and the NMS calculation is completed.
  • FIG. 5 is a schematic structural diagram of an apparatus for updating an alternative frame according to an embodiment of the present application.
  • the candidate box updating apparatus 300 may include an overlap degree acquiring module 301 and an updating module 302.
  • the overlap obtaining module 301 is configured to obtain the degree of overlap between the first candidate frame and the plurality of second candidate frames according to the location information and the location vector of the first candidate frame.
  • the first candidate box is an candidate box with the highest score of the classifier in the set of candidate boxes.
  • the second alternative box is an alternative box in the set of candidate boxes other than the first candidate box.
  • the location vector includes location information of a plurality of second candidate frames.
  • the updating module 302 is configured to compare the first candidate box with the overlap degree and the overlap degree threshold of the plurality of second candidate boxes to obtain an updated set of candidate frames.
  • the foregoing update module 302 is specifically configured to determine that the second candidate frame remains in the updated state when the overlap degree of the second candidate frame and the first candidate frame in the multiple second candidate frames is less than or equal to the overlap degree threshold. In the set of candidate boxes; determining that the second candidate box does not remain in the updated candidate when the degree of overlap of the second candidate box and the first candidate box in the plurality of second candidate boxes is greater than the overlap threshold In the box collection.
  • the overlap acquiring module 301 is further configured to use the location information of the first candidate frame in the updated candidate set of frames and the updated location vector to obtain the updated candidate set.
  • the updated location vector includes location information of a plurality of second candidate frames in the updated set of candidate frames.
  • the update module 302 is further configured to compare the overlap degree and the overlap threshold of the first candidate block in the updated set of candidate blocks with the plurality of second candidate blocks in the updated set of candidate blocks to obtain again The updated set of candidate boxes until the overlap degree of the first candidate frame in the updated candidate set of the block obtained by the overlap obtaining module 301 and the second candidate frame in the updated set of candidate blocks is greater than
  • the overlap threshold the update module 302 is further configured to stop updating the set of candidate boxes, and to use the first candidate box in the set of candidate boxes and the first candidate box in the set of candidate boxes after each update. , as the target recognition box.
  • the set of candidate boxes contains a plurality of candidate boxes for identifying faces in the image.
  • the set of candidate boxes contains a plurality of candidate boxes for identifying the target objects in the image.
  • FIG. 6 is a schematic structural diagram of hardware of an alternative frame updating device 400 in the embodiment of the present application.
  • the alternate frame update device 400 includes a memory 401, a processor 402, and a computer program stored on the memory 401 and executable on the processor 402.
  • the processor 402 described above can include a central processing unit (CPU), or an specific integrated circuit (ASIC), or can be configured to implement one or more integrated circuits of embodiments of the present application.
  • CPU central processing unit
  • ASIC specific integrated circuit
  • Memory 401 can include mass storage for data or instructions.
  • the memory 401 may comprise a HDD, a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a magnetic tape or a universal serial bus (USB) drive, or a combination of two or more of these.
  • Memory 401 may include removable or non-removable (or fixed) media, where appropriate.
  • Memory 401 may be internal or external to data query device 400, where appropriate.
  • memory 401 is a non-volatile solid state memory.
  • memory 401 includes a read only memory (ROM).
  • the ROM may be a mask programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM) or flash memory or A combination of two or more of these.
  • PROM programmable ROM
  • EPROM erasable PROM
  • EEPROM electrically erasable PROM
  • EAROM electrically rewritable ROM
  • flash memory or A combination of two or more of these.
  • the processor 402 runs a program corresponding to the executable program code by reading the executable program code stored in the memory 401 for performing the alternate frame update method in the various embodiments described above and the NMS method including the candidate frame update. .
  • the alternate frame update device 400 can also include a communication interface 403 and a bus 404. As shown in FIG. 6, the memory 401, the processor 402, and the communication interface 403 are connected by the bus 404 and complete communication with each other.
  • the communication interface 403 is mainly used to implement communication between modules, devices, units and/or devices in the embodiments of the present application. Communication interface 403 is also accessible to input devices and/or output devices.
  • Bus 404 includes hardware, software, or both, coupling components of data querying device 400 to each other.
  • bus 404 may include an accelerated graphics port (AGP) or other graphics bus, an enhanced industry standard architecture (EISA) bus, a front side bus (FSB), a super transfer (HT) interconnect, an industry standard architecture (ISA) Bus, Infinite Bandwidth Interconnect, Low Pin Count (LPC) Bus, Memory Bus, Micro Channel Architecture (MCA) Bus, Peripheral Component Interconnect (PCI) Bus, PCI-Express (PCI-X) Bus, Serial Advanced Technical Attachment (SATA) bus, Video Electronics Standards Association Local (VLB) bus or other suitable bus or a combination of two or more of these.
  • Bus 404 may include one or more buses, where appropriate. Although a particular bus is described and illustrated in this application, the present application contemplates any suitable bus or interconnect.
  • An embodiment of the present application further provides a storage medium, where the program is stored, and when the program is executed by the processor, the candidate box update method in the foregoing embodiments and the NMS method including the candidate box update may be implemented.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Image Analysis (AREA)

Abstract

La présente invention concerne un procédé et un appareil de mise à jour de boîte candidate, se rapportant au domaine des ordinateurs. Le procédé de mise à jour de boîte candidate comprend les étapes suivantes consistant : sur la base d'informations de position et d'un vecteur de position d'une première boîte candidate, à obtenir le degré de chevauchement de la première boîte candidate et d'une pluralité de secondes boîtes candidates, la première boîte candidate étant une boîte candidate ayant la note la plus élevée d'un classificateur dans un ensemble de boîtes candidates, les secondes boîtes candidates étant les boîtes candidates dans l'ensemble de boîtes candidates autres que la première boîte candidate, et le vecteur de position comprenant les informations de position de la pluralité de secondes boîtes candidates ; et à comparer le degré de chevauchement de la première boîte candidate et de la pluralité de secondes boîtes candidates avec un seuil de degré de chevauchement pour obtenir un ensemble de boîtes candidates mises à jour. À l'aide de la solution technique de la présente invention, il est possible d'améliorer l'efficacité de mise à jour des boîtes candidates et ainsi d'augmenter l'efficacité de fonctionnement de la reconnaissance et de la détection d'image.
PCT/CN2019/077397 2018-04-19 2019-03-08 Procédé et appareil de mise à jour de boîte candidate Ceased WO2019201029A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810355286.3A CN110390344B (zh) 2018-04-19 2018-04-19 备选框更新方法及装置
CN201810355286.3 2018-04-19

Publications (1)

Publication Number Publication Date
WO2019201029A1 true WO2019201029A1 (fr) 2019-10-24

Family

ID=68240408

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/077397 Ceased WO2019201029A1 (fr) 2018-04-19 2019-03-08 Procédé et appareil de mise à jour de boîte candidate

Country Status (2)

Country Link
CN (1) CN110390344B (fr)
WO (1) WO2019201029A1 (fr)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111783655A (zh) * 2020-06-30 2020-10-16 Oppo广东移动通信有限公司 一种图像处理方法、装置、电子设备和存储介质
CN114611666A (zh) * 2022-03-08 2022-06-10 安谋科技（中国）有限公司 一种nms函数的量化方法、电子设备及介质

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111145215B (zh) * 2019-12-25 2023-09-05 北京迈格威科技有限公司 一种目标跟踪方法及装置
CN112817881A (zh) * 2021-02-26 2021-05-18 上海阵量智能科技有限公司 信息处理方法、装置、设备及存储介质

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106529398A (zh) * 2016-09-23 2017-03-22 深圳云天励飞技术有限公司 基于级联结构的快速精准人脸检测方法
US20180060649A1 (en) * 2016-08-30 2018-03-01 Irida Labs S.A. Fast, embedded, hybrid video face recognition system
CN107808143A (zh) * 2017-11-10 2018-03-16 西安电子科技大学 基于计算机视觉的动态手势识别方法

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2015006791A1 (fr) * 2013-07-18 2015-01-22 A.Tron3D Gmbh Combinaison de cartes de profondeur de différents procédés d'acquisition
US11120478B2 (en) * 2015-01-12 2021-09-14 Ebay Inc. Joint-based item recognition
CN104573731B (zh) * 2015-02-06 2018-03-23 厦门大学 基于卷积神经网络的快速目标检测方法
US20170330059A1 (en) * 2016-05-11 2017-11-16 Xerox Corporation Joint object and object part detection using web supervision
CN106960178B (zh) * 2017-02-23 2020-02-07 中国科学院自动化研究所 绝缘子识别模型的训练方法以及绝缘子的识别与定位方法
CN107301383B (zh) * 2017-06-07 2020-11-24 华南理工大学 一种基于Fast R-CNN的路面交通标志识别方法
CN107577990B (zh) * 2017-08-09 2020-02-18 武汉世纪金桥安全技术有限公司 一种基于gpu加速检索的大规模人脸识别方法
CN107657224B (zh) * 2017-09-19 2019-10-11 武汉大学 一种基于部件的多层并行网络sar图像飞机目标检测方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180060649A1 (en) * 2016-08-30 2018-03-01 Irida Labs S.A. Fast, embedded, hybrid video face recognition system
CN106529398A (zh) * 2016-09-23 2017-03-22 深圳云天励飞技术有限公司 基于级联结构的快速精准人脸检测方法
CN107808143A (zh) * 2017-11-10 2018-03-16 西安电子科技大学 基于计算机视觉的动态手势识别方法

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111783655A (zh) * 2020-06-30 2020-10-16 Oppo广东移动通信有限公司 一种图像处理方法、装置、电子设备和存储介质
CN114611666A (zh) * 2022-03-08 2022-06-10 安谋科技（中国）有限公司 一种nms函数的量化方法、电子设备及介质
CN114611666B (zh) * 2022-03-08 2024-05-31 安谋科技（中国）有限公司 一种nms函数的量化方法、电子设备及介质

Also Published As

Publication number Publication date
CN110390344A (zh) 2019-10-29
CN110390344B (zh) 2021-10-26

Similar Documents

Publication Publication Date Title
US11734954B2 (en) Face recognition method, device and electronic equipment, and computer non-volatile readable storage medium
JP6871314B2 (ja) 物体検出方法、装置及び記憶媒体
KR102578209B1 (ko) 영상 처리 방법과 장치
US20200005022A1 (en) Method, terminal, and storage medium for tracking facial critical area
US20240256377A1 (en) Fault diagnosis method and apparatus, electronic device, and storage medium
CN109086734B (zh) 一种对人眼图像中瞳孔图像进行定位的方法及装置
WO2019051941A1 (fr) Procédé, appareil et dispositif d'identification de type de véhicule, et support de stockage lisible par ordinateur
WO2017197620A1 (fr) Détection d'êtres humains dans des images grâce à des informations de profondeur
WO2019201029A1 (fr) Procédé et appareil de mise à jour de boîte candidate
US8693739B2 (en) Systems and methods for performing facial detection
KR101969623B1 (ko) 병렬 검출과 추적, 및/또는 그룹화된 피처 모션 시프트 추적을 이용한 얼굴 인식
CN109726678B (zh) 一种车牌识别的方法及相关装置
CN112597909A (zh) 一种用于人脸图片质量评价的方法与设备
CN111753722B (zh) 一种基于特征点类型的指纹识别方法及装置
CN110910445A (zh) 一种物件尺寸检测方法、装置、检测设备及存储介质
KR102421604B1 (ko) 이미지 처리 방법, 장치 및 전자 기기
CN114399791B (zh) 一种行人检测方法、装置、设备和介质
CN114842205B (zh) 车损检测方法、装置、设备及存储介质
CN114898155A (zh) 车辆定损方法、装置、设备及存储介质
CN115131725A (zh) 交通流统计方法、装置、设备及存储介质
WO2025025465A1 (fr) Procédé et appareil d'évaluation de qualité d'image de visage infrarouge
CN114067145B (zh) 无源分光器检测方法、装置、设备及介质
CN113537199B (zh) 图像边界框筛选方法、系统、电子装置及介质
CN105160333B (zh) 一种车型识别方法及识别装置
CN116129496A (zh) 一种图像遮挡方法、装置、计算机设备及存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19788579

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19788579

Country of ref document: EP

Kind code of ref document: A1