Image Reconstruction Device

Abstract

The present invention provides an image reconstruction device that can reduce the model size of a neural network while suppressing a decrease in the accuracy of image reconstruction. [Solution] An image reconstruction device that performs image reconstruction using a neural network, wherein the neural network is equipped with a low-quantization means for binarizing learned parameters, and each value of the binarized parameters is replaced with two values ​​extracted from the learned parameters and used for image reconstruction.

Description

Technical Field

[0001] The present invention relates to an image reconstruction apparatus that performs image reconstruction using a neural network.

Background Art

[0002] Machine learning using a neural network, which is a mathematical model mimicking the structure of the human brain, is used in various fields. For example, a neural network with a multi-layer structure such as a convolutional neural network (CNN) can perform advanced tasks using learned parameters and is known to exhibit high accuracy in image recognition and speech recognition (see Non-Patent Document 1).

[0003] CNN is mainly composed of an input layer, a convolutional layer, and a pooling layer, and the configuration and arrangement of the convolutional layer and the pooling layer are related to the performance of CNN. In image recognition by CNN, it is necessary to perform learning in advance using a learning data set. The convolutional layer extracts local feature amounts by applying a kernel having a certain size to the input image. The kernel is a weight matrix learned based on the learning data set.

Prior Art Documents

Non-Patent Documents

[0004]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] Improving the accuracy of image recognition in CNNs, such as those described in Non-Patent Document 1, requires increasing the number of convolutional and pooling layers, which is essential for multi-layering. However, this increases the number of parameters, leading to a larger model size. While using a large kernel can also improve inference accuracy, it requires a large amount of memory and high computing power, making it difficult to implement in small devices such as mobile devices and embedded systems. However, simply reducing the number of parameters or using a small kernel reduces the neural network model size, but this comes at the cost of significantly lowering inference accuracy.

[0006] This invention addresses these problems and aims to provide an image reconstruction device that can reduce the model size of the neural network while suppressing a decrease in the accuracy of image reconstruction. [Means for solving the problem]

[0007] To solve the above problems, the image reconstruction apparatus of the present invention is An image reconstruction device that performs image reconstruction using a neural network, The neural network is characterized by comprising a low-quantization means for binarizing the learned parameters, replacing each value of the binarized parameters with two values ​​extracted from the learned parameters, and using these for image reconstruction. This feature allows the low-quantization method to reduce the model size of the neural network by binarizing the parameters learned in the neural network. Furthermore, by replacing each value of the binarized parameters with two values ​​extracted from the pre-learned parameters before binarization, the decrease in the accuracy of image reconstruction can be suppressed.

[0008] The aforementioned neural network is a convolutional neural network, The learned parameters are the kernels of the convolutional layer. The two aforementioned values ​​are characterized by being the larger and smaller values ​​in the kernel array. This feature means that the sequence of the replacement kernel will be similar to the sequence of the trained kernel before binarization, thus further suppressing the decrease in the accuracy of image reconstruction.

[0009] The aforementioned large value is the maximum value in the kernel array, The aforementioned small value is characterized by being the minimum value in the kernel array. This feature allows the replacement kernel sequence to be closer to the pre-trained kernel sequence before binarization, further suppressing the decrease in image reconstruction accuracy.

[0010] The aforementioned image reconstruction recognizes an image by inferring from the input image. The aforementioned convolutional neural network is U-Net, The low-quantization means is characterized by binarizing the kernel applied to the deep layers of the U-Net. This feature allows for improved accuracy in image reconstruction.

[0011] The image reconstruction apparatus of the present invention is An image reconstruction device that performs image reconstruction using a neural network, By executing the low-quantization function incorporated into the model that constitutes the neural network, An extraction process step of extracting and saving two values ​​from the parameters learned by the neural network, A binarization process step to obtain binarized parameters by binarizing the learned parameters, A substitution process step to obtain a substituted parameter by substituting the value of the binarization parameter with the two values, The device is characterized by performing a write-back processing step, which involves writing the substituted parameters back to the trained neural network. This feature allows for the reduction of the neural network model size by binarizing the learned parameters. Furthermore, by replacing each value of the binarized parameters with two values ​​extracted from the learned parameters before binarization, and writing the replaced parameters back into the trained neural network, the decrease in image reconstruction accuracy can be suppressed. [Brief explanation of the drawing]

[0012] [Figure 1] This figure shows an overview of image recognition using a neural network in an image reconstruction apparatus according to an embodiment of the present invention. [Figure 2] This figure shows an example of the extraction process of a trained kernel using a low-quantization means, the binarization process, and the replacement process of the binarized kernel values ​​in the embodiment. [Figure 3] This figure illustrates the image augmentation of the dataset (CamVid) used in the training phase of U-Net in Example 1. [Figure 4] This diagram shows the architecture of U-Net in Example 1. [Figure 5] (a) is a figure showing 16 trained kernels with different weight coefficients in Example 1, (b) is a figure showing the binarized kernels obtained by binarizing the trained kernels in (a) using error diffusion or simple thresholding, and (c) is a figure showing the replaced kernels obtained by replacing each value of the binarized kernel in (b) with the maximum and minimum values ​​in the array of trained kernels. [Figure 6](a) shows the inference result (output) by U-Net when using the replaced kernel to which simple threshold processing in Example 1 is applied, and (b) shows the inference result by U-Net when using the replaced kernel to which error diffusion method is similarly applied. In FIGS. 6(a) and (b), the input image (Picture), the ground truth, and the inference result (Prediction) are shown in this order from the left. [Figure 7] It is a figure showing the model of the three-layer CNN in Example 2. [Figure 8] (a) is a figure showing ten learned kernels with different weight coefficients in Example 2, and (b) is a figure showing the binary kernels obtained by binarizing the learned kernels in (a) by the error diffusion method. [Figure 9] It is a figure showing the outline of single pixel imaging in Example 3. [Figure 10] (a) is a figure showing the learned mask pattern (mask pattern before binarization) in Example 3, (b) is a figure showing the replaced mask pattern in which the values of the binary mask pattern obtained by binarizing the learned mask pattern in (a) by simple threshold processing are replaced with the maximum value and the minimum value in the array of the learned mask pattern, and (c) is a figure showing the replaced mask pattern in which the values of the binary mask pattern obtained by binarizing the learned kernel in (a) by the error diffusion method are replaced with the maximum value and the minimum value in the array of the learned mask pattern. [Figure 11] It is a figure showing the result of image reconstruction and image quality evaluation (PSNR, SSIM) when using each mask pattern in FIG. 10. [Figure 12] It is a figure showing the result of image reconstruction and image quality evaluation (PSNR, SSIM) when changing the number of mask patterns.

MODE FOR CARRYING OUT THE INVENTION

[0013] Embodiments of the present invention will be described below. However, the present invention can be implemented in many different forms and is not limited to the embodiments and examples shown below.

[0014] As shown in Figure 1, the image reconstruction device of this embodiment is an image recognition device that performs image recognition as image reconstruction using, for example, a neural network. The neural network is equipped with a low-quantization means that binarizes parameters learned using a dataset (hereinafter referred to as "trained parameters") during the training phase. Instead of writing the binarized trained parameters (hereinafter referred to as "binarized parameters") back to the trained neural network as is, the values ​​of the trained parameters before binarization are extracted, and in the reconstruction (inference) phase, each value of the binarized parameters is replaced with two values ​​extracted from the trained parameters to generate replaced parameters, and these replaced parameters are written back to the trained neural network. This makes it possible to reduce the model size of the neural network while suppressing a decrease in the accuracy of image reconstruction (inference). The inventors have confirmed that if the binarized parameters are written back to the trained neural network as is, the accuracy of image reconstruction will decrease.

[0015] Furthermore, in this embodiment, image reconstruction refers to transforming input data through several processes, and examples include image recognition by inference of the input image as described above, and image reconstruction in imaging such as single-pixel imaging.

[0016] Here, using a convolutional neural network (CNN) as an example, the extraction process, binarization process, and value replacement process of the binarized kernel (binarized parameters) by the low-quantization means in this embodiment will be explained with reference to Figure 2. It is preferable that the various processes by the low-quantization means in this embodiment be applied to a trained kernel with a size larger than the typical 3x3 kernel size, specifically a kernel size of 9x9 to 160x160. This allows for improved image reconstruction accuracy by expanding the receptive field and extracting wide-area features, while effectively reducing computational costs for trained kernels, which have high computational costs in convolution operations, and suppressing a decrease in image reconstruction accuracy.

[0017] As shown in Figure 2, first, the low-quantization means performs an extraction process to extract and store the maximum and minimum values ​​as two values ​​from the array of trained kernels before binarization. Next, the low-quantization means obtains a binarized kernel by performing a binarization process such that values ​​above the threshold are 1 and values ​​below the threshold are 0, for example, using the mean value in the array of trained kernels as a threshold.

[0018] Next, the low-quantization means performs a substitution process in which each value of the binarized kernel obtained by the binarization process is replaced with the maximum and minimum values ​​in the array of pre-trained kernels obtained by the extraction process. Specifically, in the substitution process, the low-quantization means replaces the value of 1 in the binarized kernel with the maximum value in the array of pre-trained kernels, and replaces the value of 0 in the binarized kernel with the minimum value in the array of pre-trained kernels, thereby obtaining a substituted kernel.

[0019] Finally, the low-quantization method performs a write-back operation, writing the replaced kernel obtained by the replacement process back to the convolutional layer of the trained neural network (see Figure 1).

[0020] In this way, the replaced kernel is binarized to the maximum and minimum values ​​in the array of the pre-trained kernel before binarization, resulting in an array similar to the pre-trained kernel before binarization. Therefore, when written back to the trained neural network, the decrease in image reconstruction accuracy can be suppressed.

[0021] Furthermore, the low-quantization means allows for the selection of a convolutional layer to which, for example, the extraction of a trained kernel by the low-quantization means, binarization, replacement of the values ​​of the binarized kernel, and writing back the replaced kernel are applied to the CNN.

[0022] While we have described an example of binarization using a low-quantization method by performing a simple thresholding process with the mean value in the kernel array as the threshold, various low-quantization methods other than simple thresholding can be used for binarization using a low-quantization method. The inventors have confirmed that the accuracy of image reconstruction can be improved by preferably using the error diffusion method.

[0023] Furthermore, the threshold set during the binarization process is not limited to the mean value in the kernel array; for example, it could be the median.

[0024] Furthermore, regarding the extraction process using the low-quantization means, the two extracted values ​​are not limited to the maximum and minimum values ​​in the array of trained kernels, but may also be, for example, the large and small values ​​in the array of trained kernels. Note that the large and small values ​​may be, for example, the maximum and minimum values ​​after removing outliers in the array of trained kernels, or they may be the average of the top multiple values ​​and the average of the bottom multiple values ​​in the array of trained kernels. In this way, by using the large and small values ​​extracted from the array of trained kernels in the substitution process for the binarized kernel, a substituted kernel having an array similar to the trained kernel before binarization can be obtained, as described above.

[0025] Furthermore, it goes without saying that the two values ​​extracted from the pre-trained kernel before binarization during the extraction process are set appropriately according to various conditions, as their effect on suppressing the decrease in image reconstruction accuracy varies depending on factors such as the type of low-quantization method used in the binarization process and the threshold.

[0026] Furthermore, in this embodiment, we have described an example in which a low-quantization means is applied to extract a trained kernel, perform binarization, replace the values ​​of the binarized kernel, and write back the replaced kernel to a CNN. However, it goes without saying that the various processes using the low-quantization means are not limited to CNNs and can also be applied to other neural networks such as recurrent neural networks (RNNs) and deep learning models such as Transformer.

[0027] In practice, the low-quantization means is, for example, a low-quantization function (a set of programs) incorporated into a model that constitutes a neural network that forms the basis of image recognition. That is, the image reconstruction device is a device that performs the following steps by executing a low-quantization function incorporated into a model that constitutes a neural network: an extraction step that extracts and stores two values ​​from trained parameters; a binarization step that obtains binarized parameters by binarizing the trained parameters; a substitution step that obtains substituted parameters by substituting the values ​​of the binarized parameters with the two values ​​extracted from the trained parameters; and a write-back step that writes the substituted parameters back to the trained neural network.

[0028] In this embodiment, it is preferable that the training using the dataset during the neural network training phase (see Figure 1) is performed as usual without dequantization. This allows the image reconstruction device to reduce the model size of the high-precision trained neural network while suppressing a decrease in the accuracy of image reconstruction. [Examples]

[0029] Here, the image recognition device used as the image reconstruction device in Example 1 of the above embodiment will be described in detail below with reference to Figures 3 to 6.

[0030] In this embodiment, neural network training is performed as usual. Furthermore, the training dataset used in the neural network training phase is CamVid, a dataset of road photos taken from an in-vehicle camera, which has been augmented by rotation, cropping, and flipping (see Figure 3). By performing image augmentation, it is possible to compensate for deficiencies and biases in the training data and increase the amount of training data, thereby improving inference accuracy.

[0031] Furthermore, in this embodiment, U-Net, a neural network designed for semantic segmentation, is used (see Figure 4). Since the configuration of U-Net is well-known, a detailed explanation of its configuration will be omitted for the sake of clarity. The execution environment for U-Net is shown in Table 1.

[0032] [Table 1]

[0033] As shown in Figure 4, in this embodiment, the trained kernels in the convolutional layers 1 (Conv9×9, ReLU), 2 (Conv15×15, ReLU), 3 (Conv21×21, ReLU), and 4 (Conv27×27, ReLU), indicated by numbered arrows in the U-Net architecture, are extracted, and the trained kernels are extracted using a low-quantization means, binarized, the values ​​of the binarized kernels are replaced, and the replaced kernels are written back.

[0034] In detail, the low-quantization method extracts and stores the maximum and minimum values ​​from an array of 16 pre-trained kernels with different weight coefficients (see Figure 5(a)) before binarization. Then, the trained kernels are binarized to 1 and 0 using error diffusion and simple thresholding. Figure 5(a) shows an example of a trained kernel in convolutional layer 3.

[0035] Figure 5(b) shows the binarized kernels obtained by binarizing the trained kernel in Figure 5(a) using error diffusion and simple thresholding, respectively. In this example, Floyd-Steinberg dithering is applied as the error diffusion method.

[0036] Furthermore, during inference, the values ​​of 1 in the binarized kernel are replaced with the maximum value in the array of trained kernels, and the values ​​of 0 in the binarized kernel are replaced with the minimum value in the array of trained kernels. This replaced kernel (see Figure 5(c)) is then written back to the trained U-Net to perform convolution. In this embodiment, the same series of processes are performed not only on convolutional layer 3, but also on the other convolutional layers 1, 2, and 4.

[0037] Here, Table 2 shows the results of a comparison of the inference results (outputs) by U-Net when binarized using simple thresholding and error diffusion, using mloU and pixel accuracy as evaluation metrics. Table 2 also shows the inference results when using a pre-trained kernel before binarization as a comparative example. Furthermore, Figure 6 shows the difference in inference results (predictions) when binarized using simple thresholding and error diffusion.

[0038] MloU is a metric used to measure how accurately a model can classify each pixel into the appropriate category. Pixel accuracy, on the other hand, is the accuracy rate for each individual pixel.

[0039] [Table 2]

[0040] As shown in Table 2, it was confirmed that the inference results obtained by binarizing using the error diffusion method showed suppression of the decrease in inference accuracy in both mloU and pixel accuracy compared to simple thresholding.

[0041] Furthermore, as shown in Figure 6(a), in the case of simple thresholding, the bright areas indicating vehicles disappear from the inference result image, but as shown in Figure 6(b), in the case of error diffusion, it was confirmed that bright areas indicating vehicles appear in the inference result image. In other words, the effectiveness of binarization using error diffusion was confirmed.

[0042] Next, we will explain the relationship between the reduction of the neural network model size and the inference accuracy in the image recognition device of this embodiment.

[0043] Table 3 shows the results of comparing the inference results after performing binarization processing using the error diffusion method on each of the convolutional layers 1-4 (see Figure 4) of the U-Net encoder section. Here, extraction processing using a low-quantization method, binarization processing, substitution processing, and write-back processing were performed on convolutional layers 1, 2, 3, and 4 (Pattern 1), convolutional layers 2, 3, and 4 (Pattern 2), convolutional layers 3 and 4 (Pattern 3), and convolutional layer 4 only (Pattern 4), respectively. Note that Comparative Example 2 in Table 3 shows the same results as when using the pre-trained kernel before binarization in Table 2, and Pattern 1 in Table 3 shows the same results as when binarized using the error diffusion method in Table 2.

[0044] Furthermore, Comparative Example 1 in Table 3 shows the case using a small 3x3 kernel. In Comparative Example 2, the larger kernel size compared to Comparative Example 1 results in improved inference accuracy with increased mloU and pixel accuracy, but the model size has theoretically increased by approximately 40 times. Note that the model size reduction rate in Table 3 is based on Comparative Example 2.

[0045] [Table 3]

[0046] As shown in Table 3, it was confirmed that the model size can be reduced by more than 80% while maintaining mloU and pixel accuracy by performing extraction, binarization, substitution, and write-back processing using low-quantization methods. Furthermore, when extraction, binarization, substitution, and write-back processing were performed on convolutional layers 3 and 4 (Pattern 3) and on convolutional layer 4 only (Pattern 4), respectively, it was confirmed that mloU and pixel accuracy improved compared to Comparative Example 2, which used the pre-binarization trained kernel, thus improving inference accuracy.

[0047] Thus, it was confirmed that by performing extraction, binarization, substitution, and write-back processing using low-quantization methods on the intermediate layers of the U-Net architecture, i.e., the deeper layers of U-Net, it is possible to reduce the model size while improving inference accuracy. Furthermore, it was observed that the inference accuracy tends to decrease as the target of binarization, extraction, substitution, and write-back processing using low-quantization methods becomes an early convolutional layer such as convolutional layers 1 and 2. This is presumed to be because the error diffusion method is counterproductive for small images.

[0048] In this embodiment, extraction, binarization, substitution, and write-back processing were performed on the kernel of the convolutional layer in the encoder section of U-Net using a low-quantization means. However, it goes without saying that the extraction, binarization, substitution, and write-back processing may also be performed on the kernel of the convolutional layer in the decoder section of U-Net using a low-quantization means.

[0049] Furthermore, the convolutional layers to which the extraction, binarization, substitution, and write-back processes of trained kernels using the low-quantization means are applied are not limited to the convolutional layer patterns described above (see Table 3). For example, the process may be selectively applied only to convolutional layer 3 in the encoder section of U-Net, or it may be applied to convolutional layers selected from the encoder and decoder sections of U-Net, respectively. It is preferable that the layers of the neural network to which the processing by the low-quantization means is applied are determined while checking the inference accuracy, and by applying the processing to many layers constituting the neural network, the model size can be reduced efficiently. [Examples]

[0050] Next, the image recognition device used as the image reconstruction device in Embodiment 2 according to the above embodiment will be described in detail below.

[0051] In this embodiment, a 3-layer CNN is used as the neural network (see Figure 7). Convolutional layers 1 to 3 all perform conv+BN+ReLU processing, with the size of the pre-trained kernel used in convolutional layer 1 being 160x160, the size of the pre-trained kernel used in convolutional layer 2 being 65x65, and the size of the pre-trained kernel used in convolutional layer 3 being 33x33. The execution environment for the 3-layer CNN is shown in Table 4.

[0052] [Table 4]

[0053] In this embodiment, the error diffusion method is applied to all trained kernels used in the convolutional layer 1. A binarized kernel is obtained by binarizing the array of trained kernels, using the mean value of the array as a threshold, such that values ​​above the threshold are 1 and values ​​below the threshold are -1. Table 5 shows the inference results of a CNN using error diffusion method 1, which uses a substituted kernel in which the values ​​of 1 in the binarized kernel are replaced with the maximum value in the array of trained kernels before binarization, and the values ​​of -1 in the binarized kernel are replaced with the minimum value in the array of trained kernels before binarization, and error diffusion method 2, which uses the binarized kernel as is without any substitution. As a comparative example, Table 5 also shows the inference results when using the trained kernel before binarization.

[0054] [Table 5]

[0055] As shown in Table 5, the accuracy of the CNN with error diffusion method 2 decreased significantly, but the CNN with error diffusion method 1 was found to have a significantly reduced model size without sacrificing inference accuracy, and the CPU inference time was also significantly reduced as a result of the reduced model size.

[0056] Furthermore, results from a CNN using error diffusion method 2 confirmed that directly writing the binarization parameters back into the trained neural network degrades the inference accuracy. [Examples]

[0057] Next, the imaging apparatus used as the image reconstruction apparatus in Example 3 of the above embodiment will be described in detail below.

[0058] In this embodiment, in an imaging device using single-pixel imaging with a single-pixel sensor (see Figure 9), the mask pattern (illumination pattern) that determines the image quality and resolution of the reconstructed image is optimized by deep learning. In a single-pixel imaging device, a binary optical modulator is required to project the mask pattern onto the object at high speed, so the mask pattern optimized by deep learning needs to be binaryized (binarized).

[0059] In this embodiment, the neural network is trained as usual. The training dataset used in the neural network training phase is MNIST with an image size of 128 x 128 pixels. The number of mask patterns is 128.

[0060] Furthermore, in this embodiment, a convolutional autoencoder is used as the neural network. Since the configuration of a convolutional autoencoder is well known, a detailed explanation of its configuration will be omitted for the sake of clarity.

[0061] In this embodiment, the low-quantization means extracts and stores the maximum and minimum values ​​from the array of mask patterns before binarization for a mask pattern (kernel), which is a trained parameter optimized by a convolutional autoencoder. Then, a binarized mask pattern is obtained by binarizing it to 1s and 0s using simple thresholding and miscalculation diffusion (Floyd-Steinberg dithering). During image reconstruction, the values ​​of 1 in the binarized mask pattern are replaced with the maximum value in the array of trained mask patterns, and the values ​​of 0 in the binarized mask pattern are replaced with the minimum value in the array of trained mask patterns. The resulting replaced mask pattern (see Figure 10) is written back to the trained neural network to perform convolution. The result of image reconstruction is shown in Figure 11.

[0062] As shown in Figure 11, the reconstructed image using the replaced mask pattern obtained by binarization using the miscalculation diffusion method was of higher quality compared to the reconstructed image using the replaced mask pattern obtained by binarization using simple thresholding, confirming that the decrease in image reconstruction accuracy can be suppressed. However, even when binarization was performed using the error diffusion method, the image reconstruction accuracy decreased when the binarized mask pattern was written back directly to the trained neural network.

[0063] Next, as a training dataset used in the neural network training phase, the STL-10 image, originally with an image size of 96x96px, was resized to 128x128px and converted from color to grayscale. Figure 12 shows the results when the number of mask patterns was varied from 128 to 2048. The various processing methods using the low-quantization technique were performed in the same way as described above when using MNIST.

[0064] As shown in Figure 12, compared to reconstructed images using a replaced mask pattern that has been binarized by simple thresholding, the reconstructed images using a replaced mask pattern that has been binarized by miscalculation diffusion are of higher quality, confirming that the decrease in the accuracy of image reconstruction can be suppressed.

[0065] Furthermore, even when using a replaced mask pattern that has been binarized by simple thresholding, the decrease in image reconstruction accuracy tends to be suppressed when the number of mask patterns increases to 2048. Therefore, it is presumed that by further increasing the number of mask patterns, the decrease in image reconstruction accuracy can be suppressed even when reconstructed images are made using a replaced mask pattern that has been binarized by simple thresholding.

[0066] Thus, it has been confirmed that the image reconstruction apparatus of the present invention can be used not only for image recognition and image identification as described in Examples 1 and 2, but also for the binary conversion of irradiation patterns in single-pixel imaging using deep learning, and that it can suppress the decrease in accuracy of image reconstruction.

[0067] Although embodiments of the present invention have been described above with reference to the drawings, the specific configurations are not limited to these embodiments, and any changes or additions that do not depart from the spirit of the present invention are also included. [Industrial applicability]

[0068] The present invention provides an image reconstruction device that has industrial applicability as an image reconstruction device that can reduce the model size of the neural network while suppressing a decrease in the accuracy of image reconstruction. This is achieved by replacing each value of the binarized parameters with two values ​​extracted from the learned parameters and using them for image reconstruction. Furthermore, the image recognition device as an image reconstruction device of the present invention can reduce computational costs and speed up inference due to the reduction in model size, and is expected to be applied to autonomous driving technology and edge devices. In addition, the image reconstruction device of the present invention can be applied to fields other than image recognition and image identification, for example, it can be applied to the binarization of irradiation patterns in single-pixel imaging using deep learning.

Claims

1. An image reconstruction device that performs image reconstruction using a neural network, The image reconstruction apparatus is characterized in that the neural network comprises a low-quantization means for binarizing learned parameters, and each value of the binarized parameters is replaced with two values ​​extracted from the learned parameters and used for image reconstruction.

2. The aforementioned neural network is a convolutional neural network, The learned parameters are the kernels of the convolutional layer. The image reconstruction apparatus according to claim 1, characterized in that the two values ​​are a larger value and a smaller value in the kernel array.

3. The aforementioned large value is the maximum value in the kernel array, The image reconstruction apparatus according to claim 2, characterized in that the small value is the minimum value in the kernel array.

4. The aforementioned image reconstruction recognizes an image by inferring from the input image. The aforementioned convolutional neural network is U-Net, The image reconstruction apparatus according to claim 2 or 3, characterized in that the low-quantization means binarizes the kernel applied to the deep layers of the U-Net.

5. An image reconstruction device that performs image reconstruction using a neural network, By executing the low-quantization function incorporated into the model that constitutes the neural network, An extraction process step of extracting and saving two values ​​from the parameters learned by the neural network, A binarization process step to obtain binarized parameters by binarizing the learned parameters, A substitution process step to obtain a substituted parameter by substituting the value of the binarization parameter with the two values, An image reconstruction apparatus characterized by being an apparatus that performs a write-back processing step of writing the substituted parameters back to a trained neural network.

Images