JPH03214369A - Shape storage system using neurocomputer - Google Patents

Abstract

PURPOSE:To curtail the data quantity by generating intersection information to a light beam parameter by coupling weight of a neural network, and synthesizing a three-dimensional image, based on this intersection information. CONSTITUTION:The system is provided with a shape strage part 1 using a neurocomputer, a generating part 2 for generating a shape of an object in a form of intersection information, and a three-dimensional video synthesizing part 3 for synthesizing a three-dimensional image, based on the intersection information. In such a state, based on a result of intersection calculation to an object to plural pieces of light beams inputted in advance, intersection information to a light beam parameter inputted newly is generated by coupling weight of a neural network stored in the storage part 1, and based on this intersection information, a three dimensional image is synthesized. In such a way, a massive data quantity required for realizing an image compilation is curtailed (compressed).

Description

[Detailed Description of the Invention] [Table of Contents] Overview Industrial Application Fields Prior Art and Problems to be Solved by the Invention Means for Solving the Problems Action Examples Effects of the Invention [Summary] Ray tracing in computer graphics Regarding the shape memory method used in processing, we are focusing on the calculation of invisible parts necessary for three-dimensional shape memory. The purpose of this method is to reduce (compress) the enormous amount of data required to perform visual calculations from all directions, that is, image compilation. A shape memory unit using a neurocomputer that globally stores representative ray parameters passing through grid points on each surface of Hounding Honks in the form of intersection information, and a given arbitrary ray parameter. , a generation unit using a neurocomputer that generates the shape of the object in the form of the intersection information based on the intersection information stored in the shape memory unit, and for the given arbitrary ray parameters, A 3D image synthesis unit is provided that synthesizes a 3D image based on the intersection information for arbitrary light rays output from the generation unit, and the calculation result of the intersection with the object for a plurality of input rays is provided in advance. (ray parameters), intersection information for the newly input ray parameters is generated using the neural network connection weights stored in the shape memory unit, and based on the generated intersection information, 3D Configure to perform video compositing.

[Industrial application field]

The present invention relates to a shape memory method in ray tracing processing in computer graphics.

Generally, a shape that has been modeled can be used as a component in another scene without changing its shape.

However, conventional image generation methods, including ray tracing, are methods that perform calculations from scratch.

However, as mentioned above, if you can memorize the shape of an object with a fixed definition as it is and use the memorized information to synthesize images from different scenes three-dimensionally, it will be a very effective three-dimensional A video synthesis method can be established.

The requirements for 3D image synthesis include: ■ Hidden surface processing that allows objects in the foreground to hide objects behind them ■ Expression of reflections, transmission, and shadows ■ It is necessary to see things differently depending on the viewpoint .

So-called holography has been the only conventional method for fixing the three-dimensional shape. This is to store object information as an interference pattern, including not only amplitude but also phase information.

The present invention attempts to perform the three-dimensional shape memory within the framework of a ray tracing method, unlike the holography method described above.

Since 2D images do not have depth information, for example,
If you change the viewpoint, the image will not be correct. In order to overcome this, it is necessary to perform three-dimensional shape memory, and to perform calculations for invisible parts and visible calculations from all directions.

This calculation cannot essentially be omitted. This is because the direction from which the object is viewed is unknown at the time of calculation.

Here, this visibility calculation from all directions is performed using image compilation.
ion). This can be thought of in the same way as compilation in the normal sense.

The compilation time does require a long time.

However, once compiled, video can be generated at a much lower cost than the processing time of an inkpreter (traditional video generation).

If this image compilation can be realized, it will become possible to synthesize three-dimensional images using a method different from holography.

[Prior art and problems to be solved by the invention] FIG. 3 is a diagram illustrating conventional image compilation.

As mentioned above, in order to perform image compilation, visible points for light rays in all directions must be stored for points on the object.

If we estimate the amount of data required for this, for example,
Assuming that the surface is circumscribed by a bounding box consisting of a 512 x 512 grid, and that a ray with a resolution of "1 degree" is stored for each grid point, the visible point data (T
) is 4-height floating point data, and the normal is 4-height x 3 directions - 12-height floating point data, then
For each side, 512 x 512 x 360 x 180
A data amount of X16 height #270G height is required.

Therefore, there is a problem in that it is actually difficult to memorize the shape of an object based on the result of the image compilation for the three-dimensional image synthesis.

In view of the above-mentioned drawbacks of the conventional technology, the present invention has been developed to perform shape memory in ray tracing processing using Conbuque Graphicnox. Visual computation of the image, i.e.
The purpose is to provide a shape memory method that can reduce (compress) the huge amount of data required to implement a compiler.

(Means for Solving the Problems) Fig. 1 is a diagram explaining the principle of the present invention, in which (a) shows the system configuration, and (b) shows an example of the configuration of the neural/distortion work learning device and generation device. (c) shows the input/output characteristics of the neural network, and (d) shows an example of the structure of the neural network.

The above problems are solved by a shape memory method configured as follows.

A shape memory method for ray tracing processing in computer graphics, in which a circumscribed bounding box surrounding an object is set, and representative ray parameters passing through grid points on each surface of the set bounding box are stored as intersection information. A shape memory unit 1 using a neurocomputer that stores globally in the form of a given arbitrary ray parameter, and the shape memory unit 1
a generating section 2 using a neurocomputer that generates the shape of the object in the form of the intersection information based on the intersection information stored in the above; A 3D image synthesis unit 3 is provided which synthesizes a 3D image based on intersection information for arbitrary light rays output from the 3D image synthesis unit 3, and a 3D image synthesis unit 3 is provided to synthesize a 3D image based on intersection information for arbitrary rays output from the 3D image synthesis unit 3. Then, intersection information for the newly input ray parameters is generated using the neural network connection weights stored in the storage unit 1, and a three-dimensional image is synthesized based on the generated intersection information. Configure.

[Effect]

That is, the present invention focuses on the fact that one of the characteristics of a neurocomputer is that it has the ability to learn at discrete points and the ability to interpolate for points that have not been learned and output a generated result. This characteristic is used to store the shape of objects generated by computer graphics.

That is, the memory representation of this method uses distributed representation using a neural network to efficiently store a huge amount of information.

Furthermore, the intersection information generation unit and the three-dimensional image synthesis unit use the interpolation function of the neurocomputer to generate intersection information ( Specifically, the three-dimensional images are synthesized using the generation results of visible points and normal hectares.

The intersection information (recognition pattern) learned through shape memory processing by the neurocomputer is information unique to the object,
For example, it can be registered in a data database and used for three-dimensional image synthesis.

Hereinafter, "neurocomputer learning and recognition device" The neural neural network learning device 20 consists of a learning pattern holding section 201, a neural neural network execution section 200, and a weight updating section 202. The learning pattern holding unit 201 stores input patterns (ray parameters) and . It holds the desired output pattern (cross information: teacher signal) for it.

The neural network that can be used with the Hanokubu mouth pagination method is a multilayer network.

Each layer is composed of many unitots, and the units in each layer are connected to each other with a certain weight.

By providing a set of this Z note work C input pattern and a desired output pattern (teacher signal), the weight of the new note work can be learned.

Learning proceeds as follows. A certain input pattern is given to a neural network and an output is obtained. If the output is incorrect, the correct (desired) output value is taught to the neural network. Then, the neural net,
The network adjusts the internal structure (strength and weight of connections) of the neural network so that the difference between the correct output and the actual output value is minimized. By repeating this many times, the neural network automatically learns weights that satisfy a certain input-output relationship. This learning algorithm is called the hack propagation method.

When using the neural neowork trained in this way, it returns the correct output (intersection information) taught for input patterns (ray parameters) that have been trained, but it also learns for input patterns (ray parameters) that have not been trained. It is possible to return an output pattern (intersection information) that is an interpolated input/output pattern. This is a major feature of neural networks.

The neural network learning device (see first [F(b)]) that performs this learning will be described below.

As described above, the neural network learning device 20 includes a learning pattern holding section 201, a neural network execution section 200, and a weight updating section 202.

(1) Learning pattern holding unit Input pattern and its desired output pattern (
teacher signal).

(2) Neural network execution unit It has a multi-layer network structure. Each layer is composed of many units, and a connection weight W is defined between each unit.

Each Uninote calculates the output value of the Nenotework as shown below.

If a unit receives input from multiple units,
The total sum plus the threshold value θ of the unit becomes the input value net. {See Figure 1(c)} That is, unit U
The input neJ of i is net; =ΣW i j O
j + θ, W, ,: Weight of connection from unit U4 to unit }tJi OJ = Output of unit UJ e,: Dark value of unit U8 The output value of unit is the summation net of this input, and the activation function is Calculated by applying. When the activation function is a differentiable nonlinear function, for example, the siBmoi function, the output value 0 of the unit yJ; becomes 1.

The network used in the hack pagination method is generally a multilayer network, but here we usually
The following describes the case of a three-layer network as shown in FIG. 1(d).

The three layers are called the input layer, hidden layer, and output layer, and each layer is composed of many units. Each unit of the hidden layer is connected to all units of the input layer. Each unit of the output layer is connected to all units of the input and hidden layers. There is no connection within each layer.

Each unit in the input layer is given input data to the network.

Therefore, the output value hJ of each unit UJ in the hidden layer is: netJ - ΣwJ, d, + OB1 d,: output value h of the k-th input unit,: output value wJk of the j-th hidden unit With the unit. The weight of the connection between the j-th hidden unit, e,: the darkness value of the j-th hidden unit. Also, the output value O, of each unit in the output layer is, from the above formula, h,: the output of the j-th hidden unit. i direct 0,: output value wij' of the first output unit; weight θ of the connection between the Jth hidden unit and the i-th output unit; = lld of the i-th output unit;

(3) Weight updating section This section changes the weight of the network so that the output of the network becomes a desired output.

When a certain pattern p is given, the mean squared error E between the actual output value (op,) and the desired output value (t...i) is taken.

1 E, = Σ (tp, O pi)22' As mentioned above, in order to learn a certain pattern p, all the weights in the network are changed so as to reduce this error.

・Learning rules for the output layer (Learning rules to eliminate the above errors) Unit U in the hidden layer, unit U in the output layer. The change in weight between ΔW; J(n) is Δw== (n)=ηΣ
δail”lpj +txΔwij(n1) 11) Input layer UNINOTO U5 - output layer UNINOTO U80
The weight change ΔWik(n) between ΔWi1((n
) −77Σδ9, d. , +rxΔWi, (fi1
) Here, n: number of learning α: momentum (smoothness of the s i gmo id function above) δpr= (tp.Op=) (Op; (1
op;) )・Learning rule for hidden layer The change in weight between the input layer unit U and one hidden layer unit I-UJO, ΔWJk(n), is as follows: ΔwJk(n)=ηΣδ. JhpJ +CXΔWJk(
n1) δ, J=h pj (1 −'h pj)Σδ. i
It can be found in Wij.

The weight changes ΔW, , ΔWik. Learning is performed by calculating ΔWjk and providing it to the neural network.

For details on the process of deriving the formulas that give the output values of each unit in each input layer, hidden layer, and output layer, or the formulas for learning rules in the output layer and hidden layers, see the rPDP model "Cognitive Science and Exploration of Neuronal Networks”, D. E. Ramelhar 1-, J. L. McClelland, PDP Research Group, supervised translation by Shunichi Amari, February 27, 1989,
As I am familiar with "Sangyo Tosho Co., Ltd., First Edition", I will explain it here.
Only the results will be presented here.

(2) Connection weight storage unit (data storage) The connection weight storage unit 2l stores the weights between each unit of the neural network at any learning stage.

■ ■ Generation device The generation device 22 stores the connection weight storage unit (data storage)
By loading the connection weights of the neural network stored in 21 and giving an input pattern (ray parameters) different from the one used in the above learning, the output (intersection information) of the neural network determined by the connection weights of each unit can be obtained. calculate.

When using the non-notowork learned as described above, for input patterns that have been learned, the correct output as taught above will be returned, but for input patterns that have not been trained, it will also return the input/output pattern that was learned. It functions to return the output, ie, the pattern (intersection information) of the above interpolated output.

Since the neurocomputer functions as described above,
The shape of the object is memorized as the connection weight of the neural network, and this data is stored in the neural network. Alternatively, by loading and entering new ray parameters into the neural network,
Intersection information with objects for human-powered rays that is different from that during learning (
This method has the effect of being able to obtain the visible point normal (hectare), and to synthesize a three-dimensional image for that ray using a realistic storage means and a realistic processing time.

〔Example〕

Embodiments of the present invention will be described in detail below using one page.

It is a diagram explaining the principle of the above-mentioned first fool invention, and the second fool is a diagram illustrating the principle of the present invention.
The figures show an example of an embodiment of the present invention, in which (a) shows an example of a memorized shape and its circumscribed cube, (b) shows an example of an input/output pattern to a neurocomputer,
(c) shows the relationship between the world coordinate system and the coordinate system of the circumscribed cube, (d) shows an example of data output from the neurocomputer, and (e) schematically shows an example of the system configuration. In the shape memory unit 1, first set the Hounding Boxes circumscribing the object, and then manually generate each ray of light every N degrees passing through all the grid points on the set Hounding Boxes. The parameters of the ray input above (hounding hole, surface number of the square, lattice point, angle of incidence) are input to the neurocomputer, and the intersection information obtained from the above calculation {visible point information ( For example, scalar data (T), normal Hekku (NX, Ny.
N2) as a teacher signal, and stores the correspondence between the learned f-ray parameters and the intersection information in the database (connection weight storage unit) 21 as connection weights for the neural/fusotowork, and stores the stored connections. Means for loading weights into the neural network, inputting new ray parameters, obtaining intersection information for the input rays, and synthesizing a three-dimensional image for the rays implements the present invention. This is a necessary means. Note that the same reference numerals indicate the same objects throughout the series.

Hereinafter, the shape memory method using the neurocomputer of the present invention will be explained with reference to the first diagram and FIG.

Shape memory processing according to the present invention can be divided into a stage of memorizing the shape of an object and a stage of extracting the stored information and synthesizing a three-dimensional image.

In shape memory processing, first, in order to memorize the shape of each object three-dimensionally, as shown in figure (a), a circumscribed cube (Hounding Honks) is set, and the shape of the Hounding Bonks is Emit the required amount of sample rays passing through each grid point on each surface, calculate the intersection information between each ray and the object in advance, and calculate the intersection information between the input ray and the object with respect to the ray. relate.

That is, for each grid point on the surface of the hounding hole, visible points and normal hectares for light rays at several sampled angles (α) are determined.

As mentioned above, it has been virtually impossible with the prior art to store the necessary amount of sample ray information for each grid point.

In the present invention, storage of this correspondence information is realized by learning of a neurocomputer.

The input pattern to the neurocomputer is, for each ray (i), ■ Surface number information of Hounding Hox ■ Passing grid point information for each surface (G.'', Gy') ■ Incident angle of the ray (however,
direction cosine in the world coordinate system) (α, “.α,”, α2”), and these are collectively called ray parameters.

As a teacher signal for the input pattern, information on the intersection of each of the rays with the object, that is, (1) a scalar dike (T) from the lattice point of the hounding box to the intersection of the object, (2) the normal at the intersection Hector (NX', Ny', N
2'), the neural computer learns to reduce the difference between the output pattern of the neural computer and the teacher signal, and the learning result is stored as the connection weight of the neural computer's neural computer. ,
For example, it is stored in the data cache 21. {See figure (e)} , (b) of figure t shows the above-mentioned input/output patterns to the neurocomputer, and each input/output pattern is set to fall within the range of ``0'' to ``1''. It is standardized and used.

Next, the processing at the three-dimensional image synthesis stage will be explained.

At this stage, first, as a preprocessing, the rays on the world coordinate system are transformed into the eigencoordinate system of the corresponding Hounding Honox (circumscribed cube) as shown in FIG.

This transformation allows fixed shapes to be moved or rotated to any part of the scene.

Next, it is calculated which surface and which position of the -L Hounding Honox the light ray from the viewpoint of the object passes through.

As shown in (e) Ua, this calculation is performed, for example, on a general-purpose computer by ordinary calculation of the intersection of a ray and a cube.

Next, the above memorized connection weight information is applied to the neural
After traversing the network C, the parameters of the ray from the viewpoint obtained as described above, ie, "surface number", "passing grid point (GX', Gy')", [ray angle (αu', α ,゜,α2゛)'' as an input pattern, the visible point (T: scalar display) according to the connection weight and its normal vector (NY', Ny'', NX') are interpolated. Output as a value.

Figure (d) schematically shows the operation in which the visible point (T) and its normal hector (NX', Ny) are output from the neurocomputer that has completed learning.

A desired three-dimensional image is synthesized using the visible points and their hectare information obtained as described in Section 2 above.

Figure (e) schematically shows an example of the system configuration of the above-mentioned shape memory method.

The shape memory unit 1 outputs a ray (ray parameter) that passes through each grid point of the Hounding Bonx, and calculates the intersection point (visible point) of this with the object in the Hounding Bonx and the normal vector, that is, the intersection. The information is calculated, for example, on a general-purpose computer. At this time, for example, eight faces forming the Hounding Bonx are identified.

Using the intersection information (i.e., visible point (T) and normal vector) obtained in the above calculation as a teacher signal and the above ray parameters as an input pattern, the neurocomputer performs learning for each face of the haunting box. conduct.

The connection weights of the neural network resulting from the learning become the main body of the memory, and by storing this in the data cache 21, for example, the above storage process is completed.

The generation unit 2 first calculates the intersection (ray parameter) between the ray from the viewpoint and the bounding box, for example, on a general-purpose computer.

At this time, the ray parameters are converted from the world coordinate system to the unique coordinate system of the haunting box. This means that the same ray variation is used as when storing.

Subsequently, the neural network connection weights of the corresponding objects are loaded from the data base 21 into the neural network.

For this neural network, the generation unit 2
When the ray parameters calculated in are input, interpolated intersection information is output as described above even for rays that were not given during the storage process in the shape memory section 1.

This value, that is, the visible point (T) and the normal hector (NX'
Nyl. N. 1), a three-dimensional image synthesis unit (for example, a general-purpose computer) 3 executes a known image processing algorithm (ray tracing algorithm) to perform three-dimensional image synthesis.

In this way, the present invention is capable of storing shapes in ray tracing processing in computer graphics, but the shape that has been modeled can be used as a component in another program without changing its shape, and that Focusing on the fact that there is a learning function for disastrous points and an interpolation function for unlearned points, we first set a hounding box circumscribing the object,
When inputting rays every N degrees that pass through all grid points on the Hounding Honox, the intersection (intersection information) between each ray and the object is calculated, and the parameters of the input ray (Hounding Honox) are calculated. input the surface number lattice point, incident angle) into the neurocomputer, and input the intersection information {visible point information (for example, scalar data T), normal hector (NX, N, , N2) obtained by the above calculation into the teacher signal. As a result, the correspondence between the ray parameters and the intersection information is learned, memorized as the connection weight of the neural network, stored in the database, and the stored connection weight is loaded into the neural network. The feature is that a new ray parameter is inputted, interpolated intersection information for the input ray is obtained, and a three-dimensional image for the ray is synthesized.

〔Effect of the invention〕

As explained above in detail, the shape memory method using a neurocomputer of the present invention sets a circumscribed \rounding bonox surrounding an object to memorize the shape in ray tracing processing in computer graphics. A shape memory section using a neurocomputer that globally stores representative ray parameters passing through grid points on each surface of the set Hounding Hosox in the form of cross information, and a given arbitrary a generation unit using a neurocomputer that generates the shape of the object in the form of the intersection information based on the ray parameters of and the intersection information stored in the shape memory unit; and the given arbitrary ray parameters. 3, which synthesizes a three-dimensional image based on the intersection information for arbitrary light rays output from the generation unit.
A dimensional image synthesis unit is provided, and based on the intersection calculation results (ray parameters) of one or more input rays with an object, a new image is generated using the connection weights of the neural network stored in the storage unit. Intersection information is generated for the input ray parameters, and a three-dimensional image is synthesized based on the generated intersection information, so the shape of the object is used as the connection weight of the neural network. and store this data in the neural network. Alternatively, by loading the neural network and inputting new ray parameters to the neural network, it is possible to obtain the intersection information (visible point, normal hectare) with the object for the input ray that is different from that at the time of learning, and This has the advantage of being able to synthesize three-dimensional images using a realistic storage means and a realistic processing time.

[Brief explanation of drawings]

The first part is a diagram explaining the principle of the present invention, the second part is a diagram showing an embodiment of the present invention, and the third part is a diagram explaining conventional image compilation. In the circle, 1 is a shape memory section, 2 is a cross information generation section or generation section, 20 is a neural network learning device 21 is a connection weight storage section, or data head 22 is a generation device 200 is a neural network. An execution unit 201 is a learning pattern holding unit, 202 is a weight updating unit, and 3 is a three-dimensional video synthesis unit. (a) Diagram to explain the principle of the present invention (Part 1) (b) Diagram to explain the principle of the present invention Figure l (Part 2) (c) (d) Diagram to explain the principle of the present invention Figure 1 (Part 3) Output [PX. Py, P a. NX, Ny, N2] 11
↑ Input ↑ [Surface number, GX. Gy α,,α,. α21 (b) Diagram showing the one-real addition 11 of the present invention FIG. 2 (Part 1) Cross information tpX', py', p. ' l1 NX, Ny N2'1? Ray parameter [surface number GX' + Gy' + α8 α, ゛, α2'] G゛: Point other than the learned point (d) Figure 2 showing the real dispersion I1 of the present invention (Part 2)

Claims (1)

[Claims] A shape memory method in ray tracing processing in computer graphics, comprising: setting a circumscribed bounding box surrounding an object, and passing through a representative lattice point on each surface of the set bounding box; a shape memory unit (1) using a neurocomputer that globally stores light ray parameters in the form of cross information, a given arbitrary ray parameter, and the shape memory unit (1);
a generation unit (2) using a neurocomputer that generates the shape of the object in the form of the intersection information based on the intersection information stored in step 1); A 3D image synthesis unit (3) is provided which synthesizes a 3D image based on the intersection information for arbitrary rays output from the generation unit (2), and a Based on the intersection calculation results with the object, intersection information for the newly input ray parameters is generated using the neural network connection weights stored in the shape memory section (1), and based on the generated intersection information, A shape memory method using a neurocomputer, which is characterized by synthesizing three-dimensional images.