JPH0355675A - Stereo matching method - Google Patents

Abstract

PURPOSE:To easily specify the matching of stereo pictures by applying an adverse transmission learning method to a multi-layer neural circuit network consisting of an input layer, a hidden layer, and an output layer with use of the subject and non- subject teacher signals. CONSTITUTION:A neural circuit network contains an input layer 1, a hidden layer 2, and an output layer 3 connected to each other and produces a pulse when an input signal exceeds the threshold value. A pair of stereo picture information are read at the outset, and a subject/non-subject area is set into one of both stereo picture information as the input information. Thus each cell output is obtained. The evaluation function value is calculated from the cell output of the layer 3 and a teacher signal and the correction value of the neural multiplex rate. The correction is applied to all subject/non-subject areas and a proper neural multiplex rate. Then a corresponding candidate area is decided into the other stereo picture information and the output is obtained by neural circuit network applying the neural multiplex rate decided by the information on the decided candidate area. Based on the obtained output, the corresponding area of one of both stereo picture information is decided in accordance with the subject area of other stereo picture information. In such a constitution, the stereo pictures can be easily matched with each other.

Description

DETAILED DESCRIPTION OF THE INVENTION "Field of Industrial Application" The present invention relates to a stereo matching method, and particularly to a stereo matching method that is optimal for analyzing stereoscopic photographs such as aerial photographs.

``Prior Art'' Conventionally, when performing measurements by stereoscopically viewing stereo photographs such as aerial photographs, it was necessary to search for the same point in a pair of left and right photographs and position the two photographs.
These positions were generally performed by the naked eye using a reflective stereoscope, etc., but these tasks are complex and extremely difficult, require skill, and place a burden on the operator. It was big.

Therefore, technology has been developed to automate the matching of these stereo photographs. The area correlation method has mainly been used as a pattern recognition method for stereo matching.

``Problems to be Solved by the Invention'' However, the stereo matching method using the conventional area correlation method described above can obtain a relatively high matching accuracy rate for regions containing characteristic undulations of shading; In regions with low features, the matching accuracy rate is low, making it impossible to perform satisfactory measurements. When matching using the area correlation method was performed in a region with low contrast characteristics, the response was gradual, similar to the contrast in the search region, and it was difficult to identify the matching point9. There has been a strong desire for a stereo matching method that is accurate and easy to identify matching points. "Means for Solving the Problems" The present invention has been devised in view of the above problems, and is capable of transmitting a teaching signal indicating target/non-target to a multilayer neural network composed of an input layer, a hidden layer, and an output layer. A method for matching stereo images by executing a pack propagation learning method using
a second step of setting a target region and at least one non-target region in the stereo IiI image data of one of the pair of stereo image data and setting them as input data; and a second step of setting each cell output based on the input data. The third step is to find
Process and output Ji! a fourth step of calculating an evaluation function value based on the cell output and the teacher signal; a fifth step of calculating a neural gravity correction amount based on the evaluation function value obtained in the fourth step;
The nerve weight ratio is corrected based on the correction amount of the nerve weight ratio obtained in this fifth step, and the steps 3 to 5 are sequentially repeated for all target areas and non-target areas to determine an appropriate nerve weight ratio. a seventh step of setting a plurality of corresponding candidate regions on the stereo rain image data of the other of the pair of stereo image data, and a seventh step of setting a plurality of corresponding candidate regions on the data of the plurality of corresponding candidate regions, and determining the result of the sixth step based on the data of the plurality of corresponding candidate regions. An eighth step of obtaining the output by a neural network applying the determined neural weight factor, and determining a corresponding area in the other stereo image data according to the target area of one stereo image data from the output of this neural network. It consists of a ninth step.

Furthermore, the present invention includes a step A6 of correcting the nerve weight factor based on the nerve weight factor correction amount obtained in the fifth step, and sequentially performing the third step to the fifth step for all target areas and non-target areas. A seventh step of repeating, an A8 step of calculating the sum of the evaluation function values obtained as a result of this seventh step, and determining whether the sum of the evaluation function values is within a predetermined value, and calculating this evaluation function value. If the sum of the evaluation function values is not within the predetermined value, the third step to the A8 step are repeated again to keep the sum of the evaluation function values within the predetermined value.

Furthermore, the present invention includes an 85th step of determining and storing the nerve weight correction amount based on the evaluation function value obtained in the fourth step, and performing the 3rd to 85th steps for all target regions and unpaired ffi regions. Step B6, which is repeated sequentially, Step B7, which calculates the sum of the evaluation function values obtained as a result of this 86th step, and determines whether the sum of the evaluation function values is within a predetermined value, and calculates the evaluation function. When the sum of the values is not within the predetermined value, the nerve gravity correction amount is calculated and corrected again, and the third step to the B7 step are repeated to keep the sum of the evaluation function values within the predetermined value. You can also prepare.

Further, the present invention searches from the neural network output for an area having an output that is different from but similar to the target area of the one stereo image data, from the one stereo image data,
A C9 step of adding data to the non-target region, a CIO step of setting data from the target region of the stereo image data and the non-target region formed by the C9 step, and determining each cell output using the input data. Cll step and output N1l! a C12 step of calculating an evaluation function value based on the cell output and the teacher signal; a C13 step of calculating a nerve gravity correction amount based on the evaluation function value obtained in the 12th step;
The nerve weight ratio is corrected based on the nerve weight ratio obtained in this step C13, and the steps C11 to C13 are sequentially repeated for all target areas and non-target areas to determine an appropriate nerve weight ratio. The present invention configured as described above is a method for matching stereo images by executing a backpropagation learning method. A target gA region and at least one non-target region are set in one of the pair of read stereo image data, and the corresponding data are set as input data. Then, each siIla output is calculated from this input data, and the output Jl! I
The evaluation function value is calculated based on the cell output and the teacher signal. Further, based on this evaluation function value, a nerve weight factor correction amount is determined, and the nerve weight factor is corrected using this nerve weight factor correction amount, and calculations are repeated for all target areas and non-target areas. Next, out of a pair of stereo image data, a plurality of corresponding candidate regions are set on the other stereo image data, and based on the data of these corresponding candidate regions, a neural The neural network calculates the output, and the other corresponding region corresponding to one stereo image region can be determined from the neural network output.

Further, the present invention provides for immediately correcting the nerve weight factor using the nerve weight correction amount obtained in the process of calculating the evaluation function value based on each input data, or The nerve weight factor correction amount obtained in the process of calculating is temporarily memorized, and after all input data has been processed, calculations are performed to correct the nerve weight factor, and finally the sum of the evaluation function values of the input data is determined as a predetermined value. It can also be configured to repeat until the value is within the value. Furthermore, the present invention can search for a region that is similar to but different from the target region of stereo III image data, and performs stereo matching by adding this region to the non-target region. "Principle of the Invention" The present invention employs a neural network when performing matching of stereo image data, and further applies a pack propagation method as a learning method.

A neural network is composed of multiple nerve cells (neurons), and one neuron is IlIp.
It is composed of a m-body, dendrites (signal input part), and axons (signal output part). The axial tree (signal output part) is synaptically connected to the dendrites of other neurons, forming a network.

The learning method applied to this neural network is called the pack propagation method, and the structure of the neural network is as shown in Figure 6. It has a multilayer structure with layer 2 and output RF43. Note that although there is a bond between layers, there is no bond between units within a layer.

Since a neuron can be regarded as a multi-input, single-output nonlinear element, in other words, it can be regarded as an element having a "threshold action." That is, when the total amount of input signals is higher than the threshold value, the output pulse is turned on, and when it is lower than the threshold value, the output is turned off.

Therefore, for input signals S1, S2, S3, . . . So, the output signal net is a weighted sum of products. =
Σ WI x SI ・ ・ ・ ・ (1) It is written as L. That is, by changing the weight (W), the structure of the network 1 can be changed. Note that the weight (W) takes positive, negative, or zero values, and zero represents no connection. Also, the input/output characteristic function is s
The i gmo i d function is applied. This s i g
The mo i d function is a differentiable pseudo-linear function,
For example, what is represented by l can be adopted. The value range of this function is O to l, and as the input value becomes larger, it becomes l, and as the input value becomes smaller, it becomes O. And when the input value is O, it is 0
．． It looks like it's going to be 5.

Next, the algorithm of the pack propagation learning rule will be explained. Note that any number of intermediate layers may be used, and a network without feedback connections (interlayer connections) is assumed. Here, the intermediate 712 is (hidden layer)
In other words, it is sometimes called a hidden layer. (a) First, an input signal such as an image pattern is input to the input layer. (b) Next, from the input layer 1 to the output 113, the state change of each neuron due to the signal transmission process is calculated in sequence.9 (C) The j-th neuron of the output N3 obtained in (b) above is calculated. The output is Opj, the desired output (teacher signal) of the neuron with respect to the input signal is TpJ, and the squared error of the following equation is defined as an evaluation function and calculated. Note that the given image pattern is assumed to be p.

? Ep=■Σ(TpJOpJ) 2 (3) 2 (d) Net1・so that the evaluation function becomes the local minimum value (preferably the minimum value) (that is, so that the actual output is as close as possible to the desired output) Change the synaptic connections of the work, that is, the weight ratio.

That is, it is sufficient to change the strength of all the connections so as to reduce the output error.Here, the amount of change in the weight W1 when the image pattern p is given is determined as. If further transformed, JpWha = ηδpJOpl (5) Note that Opi is the input value from unit i to unit j, and δ1 is whether unit j is a unit or an intermediate unit 1.
In the case of an output unit, δIIJ = (tpj - Opj) f' J
(ne LpJ)t (6) and for intermediate units, δGIJ = r' J (ne tpJ)Σδpk
Wk, East... (7) Equation (7) is a recursive leap number. The above is the basic algorithm of the pack propagation method, and the learning of each synaptic connection (correction of the weighting factor) proceeds from the output layer to the input layer in the opposite direction to the signal propagation. This is the reason why it is called pack protection. In this pack propagation method, the calculation of ``W'' starts from the output layer and proceeds to the unit i in the same layer. In the intermediate unit, it cannot be calculated unless the previous stage's dW is determined.9 (Since it is recursive), therefore,
Calculations are impossible unless you go back to the input layer after R. Therefore, in the pack propagation method, learning data is input and the results are output (forward). Next, change the strength of the binding to reduce the error in this result (CtIi orientation). Then input the learning data again. By repeating these steps, JW is determined so that the error is minimized.

Here, the general formula of AW is expressed as JW. (n+1)=77δpJopJ+U JWJt
(n)・・・・・(8) n is the number of times of learning, the first term on the right side is dW, and the second term is an additional term to prevent error oscillation and accelerate convergence.

"Example" An example of the present invention will be described based on the drawings. Honji/i1
The example stereo matching device includes an image scanner, arithmetic processing means, and output means. The image scanner is for reading image data from a pair of stereo photographs, and in this embodiment, it is input to the arithmetic processing means in 512×400 pixels, black and white, and 256 gradations. The arithmetic processing means is for executing stereo matching using a neural network.The output means is for outputting the calculation results of the arithmetic processing means, and is for display devices, printers, XY apertures, etc. This applies to devices such as external storage devices. Next, the regions for stereo matching will be explained based on FIG. The target area A in the example is 7×7
It is composed of pixels. Then, a non-target fA area B is set in the vicinity of the target area A in the left image. This non-fIL
Region B is used to excite the non-target output unit, and a plurality of regions B may be set. By appropriately setting this non-target region B, strict feature delineation can be performed with even the smallest details taken into consideration, even between data with unclear differences. Note that it is desirable that the spacing between the non-target MLIBs be about half of the target area A. This is because if the interval is too large, data regarding optimal line drawing may be leaked, and if the interval is too large, a lot of meaningless data will be captured.

A pair of stereo photographs is image-converted in advance so that the Y positions at the same point are the same. Then, in the right rain image, the search area C is set so that the Y coordinate is the same as the target area A (on the so-called shrimp bora line). Next, stereo matching using the backpropagation method as a learning method will be explained based on FIGS. 1 to 4.

"First Method" The first stereo matching method will be explained based on FIG. 1. Step 1 (hereinafter referred to as Sl)
The image scanner reads a pair of stereo image data from a pair of stereo images such as an aerial photograph, and inputs the data to a calculation processing means. Next, in step 82, target area A (7×7@element) is specified in the left image. Furthermore, in S3, at least l or more non-target areas B. Set B... Then, S4 is data preprocessing, for example, the black and white image data of the target area A and non-target area B are normalized to 256 levels.9 From this normalized data, learning by pack propagation 3 is executed at 85. 9 Here, learning by pack propagation (SUBR
OUTINE BP) will be explained in detail based on Figures 2 and 3. First, in S51, an initial value of the nerve weight factor is set using a random number. Then, in S52, for the number of learning data,
Calculate repeatedly. Next, in 353, the output of each cell is calculated. In this embodiment, the above equation (2) is used to calculate this output. Then, in S54, the above (3rd
) Next, in order to minimize the evaluation function value, in step 355, the nerve gravity correction amount is calculated using the above equation (8). The learning constant η and the stability constant α are determined empirically, for example, η=0.4 and α=
It is set to 0.6. Then, in S56, the nerve weight factor is corrected based on the nerve weight factor correction amount, and the process returns to S52 to repeat the calculation. Therefore, in the actual 7IT example, the nerve weight factor is corrected for each l learning data. Furthermore, in this embodiment, S57
In step 858, the sum of the evaluation functions of each data is calculated, and in step 858 it is determined whether the sum of the evaluation functions is smaller than a predetermined threshold value.

That is, in S58, it is determined whether the sum of errors is less than or equal to a predetermined value. If the sum of the evaluation functions is not below the threshold value, the calculation is repeated again using the nerve weight factor at this point. If the sum of the evaluation functions is less than or equal to a predetermined value, the final nerve weight ratio is output in S59. As a result, (S
UBROUTINE BP) ends. That is, by performing repeated calculations for all target areas A and non-gA dependent areas B, an appropriate nerve weight ratio can be determined.

Here, a modified example of the subroutine of the pack propagation method will be explained based on FIG.

The above-mentioned subroutine calculates the evaluation fifl number for each learning data, corrects the nerve weight factor based on this evaluation function, and performs repeated calculations. On the other hand, in this embodiment, S
In step 55, the nerve gravity correction amount is calculated and stored. Furthermore, in S57, the sum of the evaluation functions of each data is calculated, and the sum (E
358 whether rrr) is smaller than the specified threshold.
If the value is larger than the threshold value, the sum of each nerve weight correction amount or the average value of each nerve weight weight correction amount is calculated in S581. Then, in S56, each nerve weight factor is corrected by the correction amount calculated in step 3581, and the calculation is continued using this nerve weight factor again.

Therefore, in this embodiment, after calculating the nerve weight correction amount for all learning data, the correction is performed all at once. And at 358,
If it is determined that the sum of the evaluation functions is smaller than the castle value, the process proceeds to S59. The other steps are the same as the subroutine shown in FIG. 2, so their explanation will be omitted.

Returning to the flowchart in FIG. 1 to continue the explanation, the subroutine of the repeated back one-bit pagination method is executed in S5, so that the target area A and the non-target area B
Appropriate nerve weight ratios will be determined for all areas. Then, in S6, a search area C for the right image is set. Here, the search area C corresponds to a set of a plurality of corresponding candidate areas. Next, in step 87, a neural network is set up using the neural weight factor determined in step S5.

Then, in S8, the 7×7@element data in the search area C is sequentially stored in the arithmetic processing means, normalized, and then input to the neural network 1 network. Record cell output.

Then, in SIO, the maximum value of the "target" silk cell output is determined, and this corresponding candidate area is determined as a corresponding point. That is, the corresponding candidate area with the maximum output is set as the corresponding point of "Target area A" in the left image, and it is determined at Sll whether or not to end stereo matching. .

Second method J Next, the second stereo matching method will be explained based on FIG. This is added to the non-target area.St to 85 are the first
Since this method is similar to the method described in , the explanation will be omitted. In S13, the search area C is not set in the right image as in the first method, but is set in the left image similarly to the target area A. Then, in step S14, a neural network is set using the neural weight factor obtained in step 85. Then, in S15, image data of a plurality of corresponding candidate areas (7×7ji elementary data) within the search area C of the left image are sequentially taken into the arithmetic processing means, and then normalized to 256 levels of density and input into the neural network. Then, in S16, the "target" cell output of each corresponding candidate region is recorded. Furthermore, in S17, the area in which the output exceeds the determined value range is selected. As a result, it is possible to search for an area that is similar to, but different from, the target area of the stereo rain image data. Then, this searched approximate region is added to the non-target region B in S18, and the first matching method described above is started. In this second matching method, before executing the first matching method, a region that approximates the target region A is searched for by autocorrelation, and this region is added to the non-target region B, so that differences in shading characteristics are It has the outstanding effect of providing clarity. In the above embodiment, a point with a high autocorrelation rate in the left image is added to the non-target area B, but it may be effective to add data near the target area A in addition to this.

In this embodiment configured as described above, corresponding points can also be displayed on the output means. In addition, during the pack propagation reaction in the right image, the highest reaction point is taken as the corresponding point, but the reaction points above a certain level are picked up and used in combination with other methods, or estimated from the disparity value to obtain the optimal solution. If you perform reverse matching from the right image to the left image, etc., you can perform stereo matching with extremely high accuracy. In this embodiment, the initial setting is performed using the random number G':- in S51, but the setting weight of the previous target point may also be used.

Furthermore, the order of the data used to determine the neural gravity using the pack propagation method may be from target area A to non-target area B, or vice versa, or from target area A to non-target area. A plurality of target areas A may be used, such as from target area B to target area A.

It may be effective to learn the non-target area B not only in the vicinity of the target area A, but also at positions slightly distant from the target area A. Then, by normalizing the learning data, s
This has the advantage that it can also be used for Had Ing and the like.

Furthermore, by employing multiple non-target regions, the difference in characteristics between the "target regions" and the "non-target regions" becomes clear, and unnecessary reactions can be removed. "Effects" The present invention configured as described above includes a first step of reading a pair of stereo image data, and a target area and at least one non-target area in one of the pair of stereo image data. a second step of setting and setting it as input data, a third step of calculating the output of each cell based on this input data, and a fourth step of calculating an evaluation function value based on the output of the output RAM cell and the teacher signal. , a fifth step of determining the amount of nerve gravity correction based on the evaluation function value obtained in this fourth step, and a correction is made to the nerve weight factor based on the nerve weight factor correction amount obtained in this fifth step, and all a sixth step of sequentially repeating the third to fifth steps for the target region and the non-target region to determine an appropriate neural weight ratio; a seventh step of setting a region, an eighth step of obtaining an output by a neural network applying the neural weight factor determined as a result of the sixth step based on the data of the plurality of corresponding candidate regions, and this neural network. The process consists of a ninth step of determining a corresponding area in the other stereo image data according to the target area of one stereo rain image data from the output, so there is little difference in shading due to learning by the pack propagation method. Since strict delineation is performed with careful attention to detail between data, it has the effect of increasing the matching accuracy rate compared to the area correlation method. Furthermore, since the matching reaction is line spectral, there is an effect that matching points can be easily identified.

Furthermore, in the present invention, after calculating the nerve weight factor correction amount, the nerve weight factor may be corrected and calculated directly (the nerve weight factor is corrected for each learning data), but it is possible to calculate the sum of the evaluation functions, Determine whether the sum of the evaluation function values is within a predetermined value, correct the nerve weight factor when the sum of the evaluation function values is not within the predetermined value, and perform repeated calculations using this nerve weight factor. It is also possible to configure it as follows (after totalizing the neural weight factor correction amount of all the learning data, the correction is made all at once).

Furthermore, the present invention can search for a region that is similar to but different from the target region of stereo image data, and performs stereo matching by adding this region to the non-target region. In this case, similar and dissimilar data can be extracted from an image in the target ffi range and learned, so an outstanding effect is that extremely highly accurate stereo matching can be expected.

[Brief explanation of drawings]

The figures show an embodiment of the present invention, and FIG. 1 is a diagram explaining the first stereo matching method, FIG. Diagram 4 for explaining a modified example of the verbalization learning method.
The figure is a diagram explaining the second stereo matching method.
The figure is a diagram for explaining regions in stereo matching, and FIG. 6 is a diagram for explaining the structure of backpropagation. A...Target area B...Non-target area C...Search area

Claims (4)

[Claims] (1) A method for matching stereo images by executing a backpropagation learning method using a teacher signal indicating target/non-target in a multilayer neural network composed of an input layer, a hidden layer, and an output layer. , a first step of reading a pair of stereo image data; a second step of setting a target area and at least one non-target area in one of the pair of stereo image data and setting it as input data; A third step calculates the output of each cell based on this input data, a fourth step calculates an evaluation function value based on the output of the output layer cells and the teacher signal, and a neural network based on the evaluation function value calculated in this fourth step. Fifth step to find the weight correction amount
The nerve weight is corrected based on the nerve weight correction amount obtained in the fifth step, and steps 3 to 5 are sequentially repeated for all target and non-target areas to obtain an appropriate nerve weight. a sixth step of determining a ratio, a seventh step of setting a plurality of corresponding candidate areas on the other stereo image data of the pair of stereo image data, and a sixth step based on data of the plurality of corresponding candidate areas. An eighth step of obtaining the output by a neural network applying the neural weight factor determined as a result of the above, and calculating a corresponding area in the other stereo image data according to the target area of one stereo image data from the output of this neural network. A stereo matching method comprising a ninth step of determining. (2) A method for matching stereo images by executing a backpropagation learning method using a teacher signal indicating target/non-target in a multilayer neural network composed of an input layer, a hidden layer, and an output layer, , a first step of reading a pair of stereo image data; a second step of setting a target area and at least one non-target area in one of the pair of stereo image data and setting it as input data; A third step calculates the output of each cell based on this input data, a fourth step calculates an evaluation function value based on the output of the output layer cells and the teacher signal, and a neural network based on the evaluation function value calculated in this fourth step. Fifth step to find the weight correction amount
a sixth step of correcting the nerve weight factor based on the nerve weight factor correction amount obtained in the fifth step; and a step of sequentially repeating the third to fifth steps for all target areas and non-target areas. 7 step, an 8th step of calculating the sum of the evaluation function values obtained as a result of this seventh step, determining whether the sum of the evaluation function values is within a predetermined value, and calculating the sum of the evaluation function values. is not within the predetermined value, a ninth step of repeating the third to eighth steps again to keep the sum of the evaluation function values within the predetermined value; a 10th step of setting a corresponding candidate region, and an 11th step of obtaining the output by a neural network applying the neural weight factor determined as a result of the 9th step based on the data of the plurality of corresponding candidate regions. and a twelfth step of determining a corresponding region in the other stereo image data according to the target region of one stereo image data from the neural network output. (3) A method for matching stereo images by executing a backpropagation learning method using a teacher signal indicating target/non-target in a multilayer neural network composed of an input layer, a hidden layer, and an output layer. , a first step of reading a pair of stereo image data; a second step of setting a target area and at least one non-target area in one of the pair of stereo image data and setting it as input data; A third step calculates the output of each cell based on this input data, a fourth step calculates an evaluation function value based on the output of the output layer cells and the teacher signal, and a neural network based on the evaluation function value calculated in this fourth step. a fifth step of determining and storing the weight correction amount; a sixth step of sequentially repeating the third to fifth steps for all target areas and non-target areas;
A seventh step of calculating the sum of the evaluation function values obtained as a result of this sixth step, and determining whether the sum of the evaluation function values is within a predetermined value, and determining whether the sum of the evaluation function values is within a predetermined value. If not, calculate the nerve weight correction amount, make the correction, and repeat the third
an eighth step of repeating the steps from step to seventh step to keep the sum of the evaluation function values within a predetermined value; and a ninth step of setting a plurality of corresponding candidate regions on the other stereo image data of the pair of stereo image data. and a tenth step of obtaining the output by a neural network applying the neural weight factor determined as a result of the eighth step based on the data of the plurality of corresponding candidate regions.
and an eleventh step of determining a corresponding region in the other stereo image data according to the target region of the one stereo image data from the neural network output. (4) A method for matching stereo images by executing a backpropagation learning method using a teacher signal indicating target/non-target in a multilayer neural network composed of an input layer, a hidden layer, and an output layer, , a first step of reading a pair of stereo image data; a second step of setting a target region and at least one non-target region in one of the pair of stereo image data and setting them as input data; A third step calculates the output of each cell based on this input data, a fourth step calculates an evaluation function value based on the output of the output layer cells and the teacher signal, and a neural network based on the evaluation function value calculated in this fourth step. A fifth step of calculating the amount of correction of the nerve weight, and a correction of the nerve weight based on the amount of correction of the nerve weight calculated in this fifth step, and the third to fifth steps for all target areas and non-target areas. a sixth step of sequentially repeating the steps to determine an appropriate neural weight ratio; a seventh step of setting a plurality of corresponding candidate regions on the one stereo image data; and a sixth step of determining a suitable neural weight ratio based on the data of the plurality of corresponding candidate regions. an eighth step of obtaining an output using a neural network applying the neural weight factor determined as a result of the process; and an output that is different from but similar to the target area of one stereo image data from the neural network output. Adding a region to the non-target region
a first step of setting data from the target region of the stereo image data and the non-target region formed in the ninth step;
0 step, an 11th step of calculating the output of each cell using the input data, a 12th step of calculating the evaluation function value based on the output of the output layer cells and the teacher signal, and the evaluation function calculated in the 12th step. 13. Calculating the nerve weight correction amount based on the value
Modify the nerve weight ratio based on the nerve weight ratio obtained in this 13th step and repeat steps 11 to 13 sequentially for all target areas and non-target areas to determine an appropriate nerve weight ratio. a 14th step of setting a plurality of corresponding candidate areas on the other stereo image data of the pair of stereo data, and a 15th step of setting a plurality of corresponding candidate areas on the other stereo image data of the pair of stereo data, and a step of determining the result of the 14th step based on the data of the plurality of corresponding candidate areas. a 16th step of obtaining the output by a neural network applying the neural weight factor, and a 16th step of determining a corresponding region in the other stereo image data according to the target region of one stereo image data from the output of this neural network. A stereo matching method consisting of 17 steps.