Image abnormal target detection method and system based on superpixel blind spot network

Abstract

The present invention relates to the technical field of image abnormal target detection, and discloses an image abnormal target detection method and system based on a superpixel blind spot network, wherein the method comprises: obtaining a hyperspectral image, performing superpixel segmentation on the hyperspectral image, and obtaining a plurality of superpixel blocks; based on the center of each superpixel block, cutting out a superpixel image of a set size from the hyperspectral image; processing each of the cut superpixel images, and obtaining an image pair of a reconstructed image and a mask image in one-to-one correspondence; inputting all the image pairs into a superpixel blind spot network, training the network, and obtaining a trained superpixel blind spot network; based on the superpixel blind spot network, detecting the hyperspectral image to be detected, and obtaining the detection result of the abnormal target. The abnormal target of the hyperspectral remote sensing image is detected, and a higher accuracy rate is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of image abnormal target detection, and in particular to an image abnormal target detection method and system based on a superpixel blind spot network. Background Art

[0002] Hyperspectral images are a three-dimensional cube data structure with two spatial dimensions and one spectral dimension, covering hundreds of continuous narrow-band spectral channels, covering visible light, near infrared, short-wave infrared, mid-infrared and ultraviolet bands. By capturing the subtle spectral differences of ground objects, hyperspectral images have unique advantages in target detection.

[0003] The anomaly detection problem in hyperspectral imaging is a binary classification problem, whose goal is to divide the image into background and abnormal targets of interest. Abnormal targets usually have characteristics such as rarity, low probability of occurrence, small area, and significant spectral difference from the background. As an unsupervised target detection method, anomaly detection in hyperspectral images faces challenges such as the diversity of abnormal target types, the difficulty in distinguishing anomalies from background, and the detection accuracy is easily affected by the scene. These problems have important research value in multiple application fields.

[0004] Existing hyperspectral anomaly detection methods can be roughly divided into three categories: statistical-based methods, representation learning-based methods, and deep learning-based methods. Traditional anomaly detection algorithms mainly focus on spectral anomalies in hyperspectral images, and are usually difficult to accurately fit the real background distribution. When calculating the mean vector and covariance matrix of the background, the algorithm is easily disturbed by abnormal pixels. Deep learning methods have received widespread attention and application in recent years due to their excellent generalization performance.

[0005] Many reconstruction models are prone to the problem of identity mapping when learning hyperspectral images, that is, some abnormal pixels may inevitably appear in the reconstructed image, making it difficult to effectively separate the anomaly from the background, thereby reducing the model performance. The original blind spot network divides the receptive field into two windows, the inner and outer windows, and only uses the pixels of the outer window to predict the central pixel, so the network cannot obtain the spectral information of the central pixel of the receptive field. When the central pixel is an abnormal pixel (that is, significantly different from its surrounding pixels), the blind spot network is difficult to accurately predict / reconstruct the central pixel, resulting in a large reconstruction error. Since the network's prediction of the blind spot only depends on the information of neighboring pixels, self-information is excluded, thereby avoiding the occurrence of identity mapping. However, the original blind spot network is insufficient in identifying large-sized abnormal targets. This is because when predicting the central abnormal pixel, the network will be affected by the adjacent pixels containing the target spectral information, resulting in a lower reconstruction error for large-sized abnormal targets. To solve this problem, superpixel segmentation technology can divide adjacent pixels into irregular pixel blocks with visual significance based on features such as similar texture, color and brightness, and its adaptive size can effectively fit the task of abnormal target detection. Summary of the invention

[0006] In order to detect abnormal targets in hyperspectral remote sensing images without prior spectral knowledge, the present invention provides an image abnormal target detection method and system based on a superpixel blind spot network, achieving a higher accuracy.

[0007] On the one hand, a method for detecting abnormal targets in images based on a superpixel blind spot network is provided, including: acquiring a hyperspectral image, performing superpixel segmentation on the hyperspectral image to obtain a plurality of superpixel blocks; based on the center of each superpixel block, cropping a superpixel image of a set size from the hyperspectral image; in each cropped superpixel image, replacing the pixels inside the superpixel block with randomly selected external pixels, and setting the pixels outside the superpixel block as background pixels to obtain a reconstructed image; in the cropped superpixel image, marking the positions of the pixels inside the superpixel block as true positions to obtain a mask image; thus, a reconstructed image is obtained. image pairs with one-to-one correspondence between the image and the mask image; all the cropped superpixel images are processed according to the steps in the previous step to obtain several image pairs; all the image pairs are used as training sets; the training set is input into the superpixel blind spot network to train the network; during training, the reconstructed image is used as the input value of the network, and the mask image is used as the true value. The difference between the true value and the network prediction value is calculated through the loss function. When the loss function value no longer decreases, the training is stopped to obtain the trained superpixel blind spot network; based on the superpixel blind spot network, the hyperspectral image to be detected is detected to obtain the detection result of the abnormal target.

[0008] On the other hand, an image abnormal target detection system based on a superpixel blind spot network is provided, including: an acquisition module, which is configured to: acquire a hyperspectral image, perform superpixel segmentation on the hyperspectral image, and obtain a plurality of superpixel blocks; based on the center of each superpixel block, crop a superpixel image of a set size from the hyperspectral image; a construction module, which is configured to: in each cropped superpixel image, replace the pixels inside the superpixel block with randomly selected external pixels, set the pixels outside the superpixel block as background pixels, and obtain a reconstructed image; in the cropped superpixel image, mark the positions of the pixels inside the superpixel block as real positions, and obtain a mask image; thus, the reconstructed image is obtained. a training module, which is configured as follows: processing all the cropped super-pixel images according to the steps in the previous step to obtain a number of image pairs; taking all the image pairs as training sets; inputting the training sets into the super-pixel blind spot network to train the network; during training, taking the reconstructed image as the input value of the network, taking the mask image as the true value, calculating the difference between the true value and the network prediction value through the loss function, and stopping the training when the loss function value no longer decreases to obtain the trained super-pixel blind spot network; a detection module, which is configured as follows: detecting the hyperspectral image to be detected based on the super-pixel blind spot network to obtain the detection result of the abnormal target.

[0009] The above technical solution has the following advantages or beneficial effects: It provides a method for detecting anomalies in hyperspectral remote sensing images based on a superpixel blind spot network and combined with an adaptive weight mechanism. This method makes up for the shortcomings of existing deep learning methods in multispectral remote sensing image enhancement tasks to a certain extent. By combining superpixel segmentation with a blind spot network, it can not only avoid the identity mapping problem, but also show superior detection effects when dealing with large-sized abnormal targets. In addition, the adaptive weight mechanism in the training process effectively reduces the interference of abnormal pixels and significantly improves the reconstruction accuracy of background images. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0011] Figure 1 This is a flow chart of the method of embodiment 1.

[0012] Figure 2 This is a schematic diagram of the internal network structure of the superpixel blind spot network of Example 1.

[0013] Figure 3 This is the image after superpixel segmentation in Example 1.

[0014] Figure 4In the first embodiment, the pixels inside the super pixel block are replaced with randomly selected external pixels.

[0015] Figure 5 Schematic diagram of a mask image of Embodiment 1.

[0016] Figure 6 This is a schematic diagram of a network output image according to Embodiment 1. DETAILED DESCRIPTION

[0017] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0018] Embodiment 1, as Figure 1 As shown, this embodiment provides an image abnormal target detection method based on a superpixel blind spot network, including: S101: acquiring a hyperspectral image, performing superpixel segmentation on the hyperspectral image, and obtaining a plurality of superpixel blocks; based on the center of each superpixel block, cropping a superpixel image of a set size from the hyperspectral image; S102: in each cropped superpixel image, replacing the pixels inside the superpixel block with randomly selected external pixels, and setting the pixels outside the superpixel block as background pixels to obtain a reconstructed image; in the cropped superpixel image, marking the positions of the pixels inside the superpixel block as real positions to obtain a mask image; thus, the reconstructed image is obtained. image and mask image corresponding to each other; S103: all the cropped super-pixel images are processed according to the steps of S102 to obtain several image pairs; all the image pairs are used as training sets; the training sets are input into the super-pixel blind spot network to train the network; during training, the reconstructed image is used as the input value of the network, the mask image is used as the true value, the difference between the true value and the network prediction value is calculated through the loss function, and when the loss function value no longer decreases, the training is stopped to obtain the trained super-pixel blind spot network; S104: based on the super-pixel blind spot network, the hyperspectral image to be detected is detected to obtain the detection result of the abnormal target.

[0019] Furthermore, the step S101: obtaining a hyperspectral image, wherein a hyperspectral image refers to image data captured at multiple wavelengths, usually covering hundreds of bands. Each pixel of such an image contains a spectral information, which can be used to analyze the composition, physical properties, etc. of a substance in detail.

[0020] Furthermore, after acquiring the hyperspectral image and before performing superpixel segmentation on the hyperspectral image, it also includes: normalizing the hyperspectral image, subtracting the minimum value from all pixel values ​​in the image to obtain corrected pixel values, so that the minimum value of the corrected pixel values ​​in the image is 0, and then dividing each corrected pixel value by the maximum value of the corrected pixel values ​​to obtain a normalized pixel value, and the normalized pixel value range is 0 to 1, which can accelerate the convergence speed during network training.

[0021] Furthermore, if Figure 3 As shown, the S101: performing super-pixel segmentation on the hyperspectral image to obtain a plurality of super-pixel blocks is to use a super-pixel algorithm (simple linear iterative cluster, SLIC) to perform super-pixel segmentation on the hyperspectral image.

[0022] It should be understood that the SLIC algorithm is used to perform superpixel segmentation on the image, and regular and uniform superpixel blocks are generated by color and distance similarity. Given the significant spectral difference between abnormal pixels and their surrounding pixels, superpixel segmentation can effectively segment abnormal areas and normal areas into different superpixel blocks. Regarding parameter setting, the number of superpixel blocks is set to one-fourth of the total number of pixels, and the compactness parameter is set to 5.

[0023] Furthermore, the S101: based on the center of each superpixel block, cutting out a superpixel image of a set size from the hyperspectral image specifically includes: first, calculating the average position of all pixels in the superpixel block; second, taking the average position of all pixels as the center, cutting out a superpixel image of a set size from the hyperspectral image. A rectangular image is taken as a super-pixel image.

[0024] Furthermore, if Figure 4 As shown, S102: in each cropped superpixel image, the pixels inside the superpixel block are replaced with randomly selected external pixels, and the pixels outside the superpixel block are set as background pixels to obtain a reconstructed image, including: in each cropped superpixel image, the edge of the superpixel block is used as the boundary, the area inside the edge of the superpixel block is called the inside of the superpixel block, and the area outside the edge of the superpixel block is called the outside of the superpixel block; each pixel inside the superpixel block is randomly replaced with a pixel outside the superpixel block, and the pixels outside the superpixel block are set as background pixels to obtain a reconstructed image.

[0025] Furthermore, if Figure 5As shown, in the cropped superpixel image, the positions of pixels inside the superpixel block are marked as real positions to obtain a mask image, including: in the cropped superpixel image, the positions of pixels inside the superpixel block are marked as True to obtain a mask image; and the positions of pixels outside the superpixel block are marked as False.

[0026] It should be understood that for each superpixel block, an image of size 19×19 pixels is cropped with it as the center. The specific method is to first calculate the average position of all pixels in the superpixel block, and then crop with this position as the center. In order to ensure that each pixel in the original image can be predicted as a blind spot, it is necessary to traverse all superpixel blocks. In the cropped image, the superpixel block is regarded as a blind spot, and the purpose is achieved by removing its spectral information. The specific operation is to replace the pixels inside the superpixel block with randomly selected external pixels. At this time, the pixels inside the superpixel block constitute the "inner window", and the pixels outside the superpixel block constitute the "outer window". In view of the small proportion of abnormal pixels in the entire image, the outer window is mainly composed of background pixels. This strategy can effectively shield the spectral information in the central area of ​​the network receptive field, prevent the network from being disturbed by abnormal pixels, and thus improve its ability to reconstruct the background. At the same time, a corresponding mask image needs to be established to mark the position of the pixels inside the superpixel block as TRUE. Finally, the cropped and replaced image is saved in pairs with the corresponding mask image.

[0027] Furthermore, S103 inputs the training set into the superpixel blind spot network to train the network, wherein the superpixel blind spot network includes: an encoder and a decoder connected in sequence; the encoder includes: a first convolution block, a first pooling layer, a second convolution block, a second pooling layer and a third convolution block connected in sequence; the decoder includes: a first upsampling layer, a fourth convolution block, a second upsampling layer, a fifth convolution block, a third upsampling layer and a sixth convolution block connected in sequence; the output end of the third convolution block is connected to the input end of the first upsampling layer; the output end of the first convolution block is jump-connected to the input end of the sixth convolution block; the output end of the second convolution block is jump-connected to the input end of the fifth convolution block; the output end of the third convolution block is jump-connected to the input end of the fourth convolution block.

[0028] Furthermore, the internal structures of the first convolution block, the second convolution block, the third convolution block, the fourth convolution block, the fifth convolution block and the sixth convolution block are consistent, and the first convolution block includes: a first convolution layer, a first batch of normalization layers, a second convolution layer, a second batch of normalization layers and a first activation function layer connected in sequence.

[0029] Furthermore, the superpixel blind spot network, the working process includes: Figure 2As shown in Figure 1, the encoder contains three convolution blocks (the first, second, and third convolution blocks), each of which consists of two layers of convolution, two layers of batch normalization, and a LeakyReLU activation function, which extract increasingly advanced features in turn.

[0030] Each convolution block is followed by a pooling layer (first and second pooling layers), which reduces the spatial dimension of the feature map through the maximum pooling operation, thereby increasing the receptive field.

[0031] As the network goes deeper, the size of the feature map gradually decreases, while the number of feature channels gradually increases, thereby extracting higher-level semantic information.

[0032] The decoder gradually restores the spatial resolution of the feature map through three upsampling layers and corresponding convolution blocks. Each upsampling layer (the first, second, and third upsampling layers) uses a deconvolution operation to enlarge the size of the feature map, and combines jump connections to directly splice the feature map from the encoder to the corresponding layer of the decoder, which not only retains rich contextual information but also enhances the ability to restore details.

[0033] The network is optimized by reconstructing the output image of the network Super pixel block area and original hyperspectral image The corresponding areas in the image are as similar as possible. The spectral information of the superpixel block in the center of the receptive field is not included, which enables the network to learn the ability to predict the central pixel based on the surrounding pixels and reconstruct the background more accurately.

[0034] Furthermore, S103 inputs the training set into the superpixel blind spot network to train the network, wherein the training process of the superpixel blind spot network includes: S103-1: inputting the training set into the superpixel blind spot network, determining whether the number of training iterations reaches the set number of weight updates, if yes, updating the training weights, if not, entering S103-2.

[0035] The weight calculation process is: first, calculate the The reconstruction error of the image pair : ;in, is the number of pixels inside the superpixel, is the mask image; is the image reconstructed by the network; is the original hyperspectral image; then, according to the reconstruction error , calculate the weight of the current image pair in the loss function : ,in, and are the maximum and minimum values ​​of the reconstruction error of all image pairs.

[0036] S103-2: Determine whether the number of training iterations reaches the set number of iterations for the overall network training or the loss function value no longer decreases. If so, it means that the training is finished. If not, return to S103-1.

[0037] It should be understood that since most of the entire picture is background pixels, the network is more inclined to learn background information in the early stages of training. As the number of network iterations increases, the degree of network overfitting deepens, and the reconstructed image is more likely to contain abnormal pixels. To this end, this patent proposes an adaptive weight. The weight acts on the loss function and is used to adjust the weight of the loss function for each image pair. Image pairs with abnormal pixels in the central superpixel block have large reconstruction errors in the early stages of training. Therefore, we reduce the training weights of image pairs with large reconstruction errors in the early stages of training to prevent the model from being interfered with by abnormal pixels.

[0038] Furthermore, in S103, during training, the reconstructed image is used as the input value of the network, the mask image is used as the true value, and the difference between the true value and the network prediction value is calculated through the loss function. When the loss function value no longer decreases, the training is stopped to obtain the trained superpixel blind spot network; wherein the loss function The specific expression is: ;in, The mask map created when constructing the blind spot image pair, is the input image, is the image output by the network, is the original hyperspectral image.

[0039] like Figure 6 As shown, the network outputs an image. Regarding the training settings, the Adam optimizer is used, the learning rate is 0.0005, and a total of about 500 rounds of learning are performed.

[0040] Further, the S104: based on the superpixel blind spot network, the hyperspectral image to be detected is detected to obtain the detection result of the abnormal target, specifically comprising: inputting the image Input to the trained superpixel blind spot network to get the generated image At this point, the network becomes a background generator, and Subtract to get the reconstruction error map, the formula is . Calculate pixel by pixel for both images The anomaly detection result is obtained by calculating the distance.

[0041] The distance calculation process includes: ;in, Reconstruct the error value for a single pixel, is the number of spectral channels.

[0042] Embodiment 2: This embodiment provides an image abnormal target detection system based on a superpixel blind spot network, including: an acquisition module, which is configured to: acquire a hyperspectral image, perform superpixel segmentation on the hyperspectral image, and obtain a plurality of superpixel blocks; based on the center of each superpixel block, crop a superpixel image of a set size from the hyperspectral image; a construction module, which is configured to: in each cropped superpixel image, replace the pixels inside the superpixel block with randomly selected external pixels, set the pixels outside the superpixel block as background pixels, and obtain a reconstructed image; in the cropped superpixel image, mark the positions of the pixels inside the superpixel block as real positions, and obtain a mask image; thus, the reconstructed image is obtained. A training module is configured as follows: all the cropped super-pixel images are processed according to the steps in the previous step to obtain a number of image pairs; all the image pairs are used as training sets; the training sets are input into the super-pixel blind spot network to train the network; during training, the reconstructed image is used as the input value of the network, the mask image is used as the true value, the difference between the true value and the network prediction value is calculated through the loss function, and when the loss function value no longer decreases, the training is stopped to obtain the trained super-pixel blind spot network; a detection module is configured as follows: based on the super-pixel blind spot network, the hyperspectral image to be detected is detected to obtain the detection result of the abnormal target.

[0043] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An image abnormal target detection method based on a superpixel blind spot network is characterized by: include: Acquire a hyperspectral image, perform super-pixel segmentation on the hyperspectral image, and obtain a number of super-pixel blocks; Based on the center of each superpixel block, a superpixel image of a set size is cropped from the hyperspectral image; The method of cutting out a superpixel image of a set size from the hyperspectral image based on the center of each superpixel block specifically includes: first, calculating the average position of all pixels in the superpixel block; second, taking the average position of all pixels in the superpixel block as the center, cutting out a superpixel image of a set size from the hyperspectral image. A rectangular image, taking the rectangular image as a super-pixel image; In each cropped super-pixel image, the pixels inside the super-pixel block are replaced with randomly selected external pixels, and the pixels outside the super-pixel block are set as background pixels to obtain a reconstructed image; The method of replacing pixels inside the superpixel block with randomly selected external pixels and setting pixels outside the superpixel block as background pixels in each cropped superpixel image to obtain a reconstructed image specifically includes: in each cropped superpixel image, taking the edge of the superpixel block as the boundary, the area inside the edge of the superpixel block is referred to as the inside of the superpixel block, and the area outside the edge of the superpixel block is referred to as the outside of the superpixel block; randomly replacing each pixel inside the superpixel block with a pixel outside the superpixel block, setting pixels outside the superpixel block as background pixels, and obtaining a reconstructed image; In the cropped super-pixel image, the positions of the pixels inside the super-pixel block are marked as the true positions to obtain a mask image; The step of marking the positions of pixels inside the superpixel block as true positions in the cropped superpixel image to obtain a mask image specifically includes: marking the positions of pixels inside the superpixel block as True in the cropped superpixel image to obtain a mask image; marking the positions of pixels outside the superpixel block as False; At this point, an image pair with one-to-one correspondence between the reconstructed image and the mask image is obtained; All the cropped superpixel images are processed according to the steps in the previous step to obtain several image pairs; all the image pairs are used as training sets; the training sets are input into the superpixel blind spot network to train the network; during training, the reconstructed image is used as the input value of the network, the mask image is used as the true value, and the difference between the true value and the network prediction value is calculated through the loss function. When the loss function value no longer decreases, the training is stopped to obtain the trained superpixel blind spot network; Based on the superpixel blind spot network, the hyperspectral image to be detected is detected to obtain the detection results of abnormal targets.

2. The method for detecting abnormal targets in images based on a superpixel blind spot network as claimed in claim 1, characterized in that: After acquiring the hyperspectral image and before performing superpixel segmentation on the hyperspectral image, the method further includes: normalizing the hyperspectral image, subtracting a minimum value from all pixel values ​​in the image to obtain a corrected pixel value, so that the minimum value of the corrected pixel value in the image is 0, and then dividing each corrected pixel value by the maximum value of the corrected pixel value to obtain a normalized pixel value, and the normalized pixel value range is 0 to 1.

3. The method for detecting abnormal targets in images based on a superpixel blind spot network as claimed in claim 1, characterized in that: The hyperspectral image is segmented into superpixels to obtain a number of superpixel blocks, and the superpixel algorithm is used to segment the hyperspectral image into superpixels.

4. The method for detecting abnormal targets in images based on a superpixel blind spot network as claimed in claim 1, wherein: Input the training set into the superpixel blind spot network to train the network, wherein the superpixel blind spot network includes: an encoder and a decoder connected in sequence; the encoder includes: a first convolution block, a first pooling layer, a second convolution block, a second pooling layer, and a third convolution block connected in sequence; the decoder includes: a first upsampling layer, a fourth convolution block, a second upsampling layer, a fifth convolution block, a third upsampling layer, and a sixth convolution block connected in sequence; the output end of the third convolution block is connected to the input end of the first upsampling layer; the output end of the first convolution block is jump-connected to the input end of the sixth convolution block; the output end of the second convolution block is jump-connected to the input end of the fifth convolution block; the output end of the third convolution block is jump-connected to the input end of the fourth convolution block; The internal structures of the first convolution block, the second convolution block, the third convolution block, the fourth convolution block, the fifth convolution block and the sixth convolution block are consistent. The first convolution block includes: a first convolution layer, a first batch of normalization layers, a second convolution layer, a second batch of normalization layers and a first activation function layer connected in sequence.

5. The method for detecting abnormal targets in images based on a superpixel blind spot network as claimed in claim 1, characterized in that: The training set is input into the superpixel blind spot network to train the network. The training process of the superpixel blind spot network includes: (1) Input the training set into the superpixel blind spot network and determine whether the number of training iterations reaches the set number of weight updates. If yes, update the training weights. If not, proceed to (2). The weight calculation process is: first, calculate the The reconstruction error of the image pair : ； in, is the number of pixels inside the superpixel, is the mask image; is the image reconstructed by the network; is the original hyperspectral image; Then, according to the reconstruction error , calculate the weight of the current image pair in the loss function : ， in, and is the maximum and minimum value of the reconstruction error of all image pairs; (2) Determine whether the number of training iterations reaches the set number of iterations for the overall network training or the loss function value no longer decreases. If so, it means that the training is over. If not, return to (1).

6. The method for detecting abnormal targets in images based on a superpixel blind spot network as claimed in claim 5, characterized in that: During training, the reconstructed image is used as the input value of the network, and the mask image is used as the true value. The difference between the true value and the network prediction value is calculated through the loss function. When the loss function value no longer decreases, the training is stopped to obtain the trained superpixel blind spot network; among them, the loss function The specific expression is: ； in, The mask map created when constructing the blind spot image pair, is the image output by the network, is the original hyperspectral image.

7. The method for detecting abnormal targets in images based on a superpixel blind spot network as claimed in claim 1, characterized in that: Based on the superpixel blind spot network, the hyperspectral image to be detected is detected to obtain the detection results of abnormal targets, including: The input image Input to the trained superpixel blind spot network to get the generated image ; The network becomes a background generator, and Subtract to get the reconstruction error map, the formula is ; Calculate pixel by pixel for the two images Distance to get anomaly detection results; The distance calculation process includes: ;in, Reconstruct the error value for a single pixel, is the number of spectral channels.

8. An image abnormal target detection system based on a superpixel blind spot network using the method of claim 1, characterized in that: include: An acquisition module is configured to: acquire a hyperspectral image, perform super-pixel segmentation on the hyperspectral image, and obtain a plurality of super-pixel blocks; Based on the center of each superpixel block, a superpixel image of a set size is cropped from the hyperspectral image; A construction module is configured to: in each cropped superpixel image, replace the pixels inside the superpixel block with randomly selected external pixels, set the pixels outside the superpixel block as background pixels, and obtain a reconstructed image; in the cropped superpixel image, mark the positions of the pixels inside the superpixel block as true positions, and obtain a mask image; thus, an image pair with a one-to-one correspondence between the reconstructed image and the mask image is obtained; The training module is configured to: process all the cropped super-pixel images according to the steps of the previous step to obtain a number of image pairs; use all the image pairs as training sets; input the training sets into the super-pixel blind spot network to train the network; During training, the reconstructed image is used as the input value of the network, and the mask image is used as the true value. The difference between the true value and the network prediction value is calculated through the loss function. When the loss function value no longer decreases, the training is stopped to obtain the trained superpixel blind spot network. The detection module is configured to: detect the hyperspectral image to be detected based on the superpixel blind spot network to obtain the detection result of the abnormal target.

Images