Progressive intelligent recognition method of small spatial targets driven by imaging prior

Abstract

The present invention relates to an imaging prior-driven, progressive, intelligent recognition method for small spatial targets, comprising the following steps: using a threshold-embedded minimum bounding rectangle algorithm to preliminarily locate a motion-blurred spatial target, obtaining a preliminarily located image; extracting degradation features from the preliminarily located image, and calculating the residual between the preliminarily located image and the degradation features to obtain a deblurred spatial target image; using a trained improved enhanced super-resolution generative adversarial network to extract features from the deblurred spatial target image, obtaining a sharpened spatial target image; and using a trained residual network to perform target recognition on the sharpened spatial target image, obtaining the spatial target type. This recognition method first preliminarily locates the spatial target, then deblurs and sharpens the located spatial target image, and finally uses the residual network for target recognition, thereby improving the accuracy of spatial target recognition in wide-format images.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and in particular relates to a progressive small-target intelligent recognition method driven by imaging priori. Background Art

[0002] The task of object detection is to find all objects of interest in an image and determine their locations and categories.

[0003] In images of small spatial targets, the proportion of the spatial targets is very small. Furthermore, due to degradation phenomena such as low resolution, high edge blur, and strong noise interference caused by factors such as camera parameters, motion patterns, and lighting environment, the spatial targets in the image are blurred. This poses a huge challenge to the detection and recognition of spatial targets. Most existing methods use direct recognition methods to detect small spatial target images, simultaneously locating, detecting, and identifying the spatial targets. However, direct recognition has low accuracy for wide-area small spatial target images. Summary of the Invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a progressive small-target intelligent recognition method driven by imaging priors. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0005] The embodiment of the present invention provides a progressive small-target intelligent recognition method driven by imaging priori, comprising the steps of:

[0006] S1. Use the minimum bounding rectangle algorithm based on threshold embedding to preliminarily locate the spatial target with motion blur and obtain a preliminary positioning image;

[0007] S2. extracting degradation features of the preliminary positioning image, and calculating the residual between the preliminary positioning image and the degradation features to obtain a deblurred spatial target image;

[0008] S3. Using the trained improved enhanced super-resolution generative adversarial network to perform feature extraction on the deblurred space target image to obtain a clear space target image;

[0009] S4. Utilize the trained residual network to perform target recognition on the cleared space target image to obtain the space target type.

[0010] In one embodiment of the present invention, step S1 includes:

[0011] The area where the space target is located is determined according to the outline of the space target, the maximum point and the minimum point of all pixel values ​​in the area are obtained, and a rectangular frame is made according to the maximum point and the minimum point to obtain the preliminary positioning image.

[0012] In one embodiment of the present invention, step S1 includes:

[0013] Determine several regions where the space target is located according to the outline of the space target, obtain the maximum value point and the minimum value point of all pixel values ​​in each of the regions, and draw rectangular frames based on the maximum value point and the minimum value point to obtain several rectangular frames;

[0014] The plurality of rectangular frames are merged according to their intersection-over-union ratio to obtain the preliminary positioning image, and the merging rule is:

[0015]

[0016] Among them, X1 BL and Y1 BL Respectively represent the horizontal and vertical coordinates of the lower left corner of the first rectangular box, X2 BL and Y2 BL They represent the horizontal and vertical coordinates of the lower left corner of the second rectangular box, IOU represents the intersection-over-union ratio of the two rectangular boxes, and threshold represents the set threshold.

[0017] In one embodiment of the present invention, step S2 includes:

[0018] Randomly cropping the preliminary positioning image to obtain a first image block and a second image block;

[0019] Performing data enhancement on both the first image block and the second image block to obtain a first enhanced image and a second enhanced image;

[0020] Using an encoder to perform feature extraction on both the first enhanced image and the second enhanced image to obtain first extracted features and second extracted features;

[0021] Inputting the first extracted features and the second extracted features into a fully connected layer, and maximizing the mutual information between the first extracted features and the second extracted features to obtain the degraded features;

[0022] The residual of the preliminary positioning image and the degradation feature is calculated to obtain a deblurred spatial target image.

[0023] In one embodiment of the present invention, the improved enhanced super-resolution generative adversarial network includes a generator and a discriminator connected in sequence;

[0024] Step S3 includes:

[0025] S31. Training the generator and the discriminator using a low-resolution image and a real high-resolution image corresponding to the low-resolution image to obtain the trained improved enhanced super-resolution generative adversarial network, wherein the low-resolution image is a deblurred spatial target image obtained by sequentially undergoing preliminary positioning and deblurring processing;

[0026] S32. Input the deblurred spatial target image into the trained improved enhanced super-resolution generative adversarial network, and use the trained generator and discriminator to perform feature extraction in sequence to obtain the cleared spatial target image.

[0027] In one embodiment of the present invention, step S31 includes:

[0028] S311, inputting the low-resolution image into the convolution layer, the rectified linear unit, the multiple dense residual blocks and the upsampling network in the generator in sequence for feature extraction to obtain a reconstructed high-resolution image;

[0029] S312, using the output of the discriminator in the previous iterative step and the reconstructed high-resolution image to calculate the adversarial loss of the generator, and using the adversarial loss of the generator to update the parameters of the generator to obtain an updated generator;

[0030] S313, inputting the real high-resolution image and the reconstructed high-resolution image into the discriminator for feature extraction to obtain the output of the discriminator; wherein the discriminator adopts a VGG network;

[0031] S314, calculating the adversarial loss of the discriminator using the output of the discriminator, the true high-resolution image, and the reconstructed high-resolution image, and training the discriminator using the adversarial loss of the discriminator to obtain an updated discriminator;

[0032] S315. Calculate the perceptual loss of the improved enhanced super-resolution generative adversarial network using the low-resolution image and the reconstructed high-resolution image during training, and optimize the improved enhanced super-resolution generative adversarial network using the perceptual loss;

[0033] S316. When the convergence condition is reached, the trained improved enhanced super-resolution generative adversarial network is obtained.

[0034] In one embodiment of the present invention, the adversarial loss of the generator is:

[0035]

[0036] in, represents the adversarial loss function of the generator, x r and x f represent the low-resolution image and the reconstructed high-resolution image, respectively. and Respectively represent the average of image data within a batch of low-resolution images and reconstructed high-resolution images, D Ra (x r , x f )=σ(C(x r )-E[C(x f )]), D Ra (x f , x r )=σ(C(x f )-E[C(x r )]), σ represents the sigmoid function, and C(·) represents the output of the discriminator in the previous iterative step.

[0037] In one embodiment of the present invention, the adversarial loss of the discriminator is:

[0038]

[0039] in, Represents the loss function of the relative average discriminator, x r and x f represent the real high-resolution image and the reconstructed high-resolution image, respectively. and Respectively represent the mean of image data within a batch of real high-resolution images and reconstructed high-resolution images, D Ra (x r , x f )=σ(C(x r )-E[C(x f )]), D Ra (x f , x r )=σ(C(x f )-E[C(x r )]), σ represents the sigmoid function, and C(·) represents the output of the relative average discriminator.

[0040] In one embodiment of the present invention, the perceptual loss includes content loss and generation loss, wherein:

[0041] The content loss is Euclidean distance loss;

[0042] The generation loss is:

[0043]

[0044] in, represents the probability that the discriminator identifies the reconstructed high-resolution image as the real high-resolution image.

[0045] In one embodiment of the present invention, the residual network includes a ResNet-50 network.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The recognition method of the present invention first performs a preliminary positioning of the space target to find out its position, then uses the imaging prior conditions to de-motion blur and sharpen the located space target image, and finally uses the residual network to perform target recognition, thereby removing the interference in the space target image and restoring the space target image with rich detail information, further improving the accuracy of space target recognition in wide-width images. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A flowchart of a method for progressive small-target intelligent recognition in space driven by imaging priors provided by an embodiment of the present invention;

[0049] Figure 2 A schematic diagram of a space target in the BUAA-SID-Share1.0 dataset provided by an embodiment of the present invention;

[0050] Figure 3 A schematic diagram of a space target after fuzzy processing provided by an embodiment of the present invention;

[0051] Figure 4 A schematic diagram of an image of a blurred space target added to a wide deep space background provided by an embodiment of the present invention;

[0052] Figure 5 It is a schematic diagram of the located space target image with motion blur;

[0053] Figure 6 It is a schematic diagram of a space target image after deblurring and sharpening operations. DETAILED DESCRIPTION

[0054] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0055] Example 1

[0056] See Figure 1 , Figure 1 The following is a flow chart of a method for progressive small-target intelligent recognition in space driven by imaging priors provided by an embodiment of the present invention. The recognition method includes the following steps:

[0057] S1. Use the minimum bounding rectangle algorithm based on threshold embedding to preliminarily locate the spatial target with motion blur and obtain a preliminary positioning image.

[0058] In this embodiment, the space target may be a small space target or a large space target.

[0059] In a specific embodiment, the area where the space target is located is determined based on the outline of the space target, the maximum and minimum points of all pixel values ​​in the area are obtained, and a rectangular frame is drawn based on the maximum and minimum points to obtain a preliminary positioning image.

[0060] Taking a small spatial target as an example, we first extract the outline of the small spatial target, determine the area where the small spatial target is located based on the outline, and then traverse all pixel values ​​in the area, calculate the maximum and minimum values, and obtain the lower left and upper right points of the rectangular box. The position of the small spatial target can be located. The formula for calculating the lower left and upper right points of the rectangular box is:

[0061]

[0062] Among them, X BL and Y BL Respectively represent the horizontal and vertical coordinates of the lower left point of the rectangle, X TR and Y TR Respectively represent the horizontal and vertical coordinates of the upper right point of the rectangle, X min and Y min Represents all horizontal coordinates {x1, x2, ..., x n} and all vertical coordinates {y1,y2,…,y n}, X max and Y max Represents all horizontal coordinates {x1, x2, ..., x n} and all vertical coordinates {y1,y2,…,y n} maximum value.

[0063] In another embodiment, due to the characteristics of the space target itself, the target pixels may be discontinuous, and in this case, multiple regions containing space target components will be obtained.

[0064] At this point, first, the several regions where the space target is located are determined based on the outline of the space target. The maximum and minimum points of all pixel values ​​in each region are obtained, and rectangular frames are drawn based on the maximum and minimum points to obtain several rectangular frames. Then, the rectangular frames are merged based on the intersection-and-union ratio of the rectangular frames to obtain a preliminary positioning image. Specifically, by artificially setting a threshold, when the intersection-and-union ratio between multiple rectangular frames is greater than the threshold, the rectangular frames are merged. The merging rule is:

[0065]

[0066] Among them, X1 BL and Y1 BL Respectively represent the horizontal and vertical coordinates of the lower left corner of the first rectangular box, X2 BL and Y2 BL They represent the horizontal and vertical coordinates of the lower left corner of the second rectangular box, IOU represents the intersection-over-union ratio of the two rectangular boxes, and threshold represents the set threshold.

[0067] It is understandable that after obtaining multiple rectangular frames, the rectangular frames are merged in pairs, and then the merged rectangular frames are merged in pairs again.

[0068] After initially locating the space target and obtaining a preliminary positioning image, a super-resolution network driven by prior knowledge of image degradation is constructed to deblur and sharpen the motion-blurred preliminary positioning image after positioning to enrich the detailed information of the space target. The specific operation steps include S2 and S3.

[0069] S2. Extract the degraded features of the preliminary positioning image and calculate the residual between the preliminary positioning image and the degraded features to obtain a deblurred spatial target image.

[0070] After initial localization using images of a class of space targets, the degradation network model is trained. Deblurring is then performed on images of similar space targets to be identified. For example, when the space targets are satellites, they are all in a space environment, and the background noise and interference in their images are essentially the same. Therefore, by training the degradation model network using space targets that are all in a space environment, the network can be used to deblur other types of space targets in space environments, demonstrating strong generalization capabilities.

[0071] First, the degradation model network is trained to obtain a trained degradation model network, which specifically includes the following steps:

[0072] 1) Set the number of training iterations.

[0073] 2) Randomly cropping a preliminary positioning image of a type of space target image to obtain a first image block and a second image block, wherein the first image block and the second image block contain different space target components.

[0074] 3) Performing data enhancement on the first image block and the second image block to obtain a first enhanced image and a second enhanced image.

[0075] Specifically, data enhancement such as translation, rotation, and scaling are performed on both the first image block and the second image block to obtain a first enhanced image and a second enhanced image.

[0076] 4) Using an encoder, extract features from the first enhanced image and the second enhanced image to obtain first extracted features and second extracted features.

[0077] Specifically, the encoder uses 6 3×3 convolutional layers and one average pooling layer. That is, the enhanced spatial target image is first passed through 6 3×3 convolutional layers to extract the feature map, and then passed through an average pooling layer for dimensionality reduction to obtain the first extracted features and the second extracted features.

[0078] 5) Inputting the first extracted features and the second extracted features into the fully connected layer, and maximizing the mutual information of the first extracted features and the second extracted features to obtain the degraded features.

[0079] 6) Repeat the above steps iteratively until the training is completed to obtain a trained degradation model network.

[0080] Then, the steps of inputting the preliminary positioning image into the trained degradation model network to extract degradation features are the same as the steps of training the degradation model network, specifically including: inputting the preliminary positioning image into the trained degradation model network, performing random cropping, data enhancement, feature extraction and maximizing mutual information on the preliminary positioning image in sequence to obtain degradation features.

[0081] Finally, the residual between the preliminary positioning image and the degraded features is calculated to obtain the deblurred spatial target image.

[0082] For images taken under the same environment, their degradation characteristics are consistent. That is, for images of different categories of spatial targets taken under the same environment, their blur kernels are the same. A low-resolution image with motion blur (i.e., a preliminary positioning image) can be obtained from a clear high-resolution image and a blur kernel. The model formula is:

[0083] I B =k(M)*I S +N

[0084] Among them, I B Represents a low-resolution image with motion blur, I S represents a clear high-resolution image, k(M) represents the blur kernel, N represents additive noise, and * represents the convolution operation.

[0085] Therefore, the deblurred spatial target image can be obtained by calculating the residual of the preliminary positioning image and the degraded features.

[0086] S3. Use the trained improved enhanced super-resolution generative adversarial network to extract features from the deblurred space target image to obtain a clear space target image.

[0087] Specifically, the improved enhanced super-resolution generative adversarial network includes a generator and a discriminator connected in sequence. The generator is used to reconstruct a high-resolution image from a low-resolution image, and the discriminator mainly calculates the discriminant loss to determine whether the input image is a high-resolution image reconstructed by the generator or the true value of the high-resolution image (i.e., the real high-resolution image).

[0088] Step S3 specifically includes the following steps:

[0089] S31. Train the generator and discriminator using the low-resolution image and the corresponding real high-resolution image to obtain a trained improved enhanced super-resolution generative adversarial network, wherein the low-resolution image is a deblurred spatial target image obtained by performing preliminary localization and deblurring processing on a similar spatial target image. In this embodiment, the similar spatial target image refers to an image with substantially the same background and interference noise as the motion-blurred spatial target.

[0090] Step S31 specifically includes:

[0091] S311. The low-resolution image is sequentially input into the convolutional layer, the rectified linear unit, multiple dense residual blocks and the upsampling network in the generator for feature extraction to obtain a reconstructed high-resolution image.

[0092] Specifically, during the training of the generator, the low-resolution image is first passed through a convolutional layer and a rectified linear unit to extract shallow features; then, the shallow features are passed through multiple dense residual blocks to extract deep features, where multiple dense residual blocks are directly connected from the previous state to all layers of the current dense residual block through dense connections; finally, the extracted deep features are input into the upsampling network. After two upsamplings, the size of the image is increased to four times the original size, and a reconstructed high-resolution image is obtained.

[0093] S312. Calculate the adversarial loss of the generator using the output of the discriminator in the previous iterative step and the reconstructed high-resolution image, and use the adversarial loss of the generator to update the parameters of the generator to obtain an updated generator.

[0094] Specifically, the discriminator uses a relative average discriminator. This discriminator estimates the probability that the true high-resolution image is relatively more realistic than the reconstructed high-resolution image. The output of the relative average discriminator, after subtracting the reconstructed high-resolution image from the true high-resolution image, should be as close to 1 as possible. The output of the relative average discriminator, after subtracting the true high-resolution image from the reconstructed high-resolution image, should be as close to 0 as possible.

[0095] The adversarial loss of the corresponding generator is:

[0096]

[0097] in, represents the adversarial loss function of the generator, x r and x f represent the low-resolution image and the reconstructed high-resolution image, respectively. and Respectively represent the average of image data within a batch of low-resolution images and reconstructed high-resolution images, D Ra (x r , x f )=σ(C(x r )-E[C(x f )]), D Ra (x f , x r )=σ(C(x f )-E[C(x r )]), σ represents the sigmoid function, and C(·) represents the output of the discriminator in the previous iterative step.

[0098] Furthermore, the adversarial loss value of the generator calculated above is added to the parameters of the generator to optimize the generator towards reconstructing high-resolution images with better quality, thereby obtaining an updated generator and finally a trained generator.

[0099] S313. Input the real high-resolution image and the reconstructed high-resolution image into a discriminator for feature extraction to obtain the output of the discriminator; wherein the discriminator adopts a VGG network.

[0100] In this embodiment, the relative average discriminator is implemented using the VGG network.

[0101] Specifically, first, the reconstructed high-resolution image and the real high-resolution image are used as input, and pass through 5 groups of VGG blocks and the maximum pooling layer in sequence. The function of the maximum pooling layer is to perform halved pooling; then feature extraction is performed through 3 dense layers and 3 fully connected layers to achieve the goal of shortening the distance between the reconstructed high-resolution image and the real high-resolution image.

[0102] S314. Calculate the adversarial loss of the discriminator using the output of the discriminator, the true high-resolution image, and the reconstructed high-resolution image, and train the discriminator using the adversarial loss of the discriminator to obtain an updated discriminator.

[0103] Specifically, the adversarial loss relative to the average discriminator is calculated, and the discriminator is trained using the adversarial loss, so that the discriminator is optimized to better distinguish between the reconstructed high-resolution image and the real high-resolution image, thereby obtaining an updated discriminator and finally a trained discriminator.

[0104] Specifically, the adversarial loss of the discriminator is:

[0105]

[0106] in, Represents the loss function of the relative average discriminator, x r and x f represent the real high-resolution image and the reconstructed high-resolution image, respectively. and Respectively represent the mean of image data within a batch of real high-resolution images and reconstructed high-resolution images, D Ra (x r , x f )=σ(C(x r )-E[C(x f )]), D Ra (x f , x r )=σ(C(x f )-E[C(x r )]), σ represents the sigmoid function, and C(·) represents the output of the relative average discriminator.

[0107] S315. Calculate the perceptual loss of the improved enhanced super-resolution generative adversarial network using the low-resolution images and the reconstructed high-resolution images during the training process, and optimize the improved enhanced super-resolution generative adversarial network using the perceptual loss.

[0108] Specifically, perceptual loss includes content loss and generation loss.

[0109] First, we calculate the content loss, which is designed as L2 loss, or Euclidean distance. L2 loss trains the network by shortening the Euclidean distance of feature maps to reconstruct a high-resolution image from the original low-resolution image.

[0110] Then calculate the generation loss, the expression of the generation loss is:

[0111]

[0112] in, represents the probability that the discriminator identifies the reconstructed high-resolution image as the real high-resolution image. The role of this loss is to make The larger the value, the better, so that when the gradient is updated The smaller.

[0113] Finally, the improved enhanced super-resolution generative adversarial network is optimized using perceptual loss to achieve sharper spatial target images. Specifically, the perceptual loss is calculated using the parameters of the generator, which can improve the texture realism of the reconstructed high-resolution image. Therefore, adding the perceptual loss to the objective function of the improved enhanced super-resolution generative adversarial network generator optimizes the generator in the improved enhanced super-resolution generative adversarial network towards reconstructing high-resolution images with less perceptual loss.

[0114] S316. When the convergence condition is reached, a trained improved enhanced super-resolution generative adversarial network is obtained.

[0115] Specifically, a convergence condition is set before training. For example, the training condition can be the number of training times or the thresholds of various losses during the training process. When the training reaches the convergence condition, a trained improved enhanced super-resolution generative adversarial network is obtained.

[0116] S32. Input the deblurred spatial target image into the trained improved enhanced super-resolution generative adversarial network, and use the trained generator and discriminator to perform feature extraction in sequence to obtain a clear spatial target image.

[0117] Specifically, the deblurred spatial target image is input into the trained generator and discriminator for feature extraction, thereby obtaining a reconstructed high-resolution image, i.e., a sharpened spatial target image. The feature extraction process does not require the calculation of various losses, and the remaining process is the same as step S31, namely, first, the low-resolution image is sequentially input into the convolutional layer, rectified linear unit, multiple dense residual blocks, and upsampling network in the generator for feature extraction to obtain a reconstructed high-resolution image; then, the true high-resolution image and the reconstructed high-resolution image are input into the discriminator for feature extraction to obtain the output of the discriminator, thereby obtaining a sharpened spatial target image.

[0118] S4. Use the trained residual network to perform target recognition on the cleared space target image to obtain the space target type.

[0119] Specifically, the residual network includes the ResNet-50 network.

[0120] First, the ResNet-50 network is trained until convergence using the sharpened spatial target image as input to obtain a trained residual network. Then, the trained residual network is used to perform target recognition on the sharpened spatial target image to obtain the spatial target type.

[0121] Specifically, training the ResNet-50 network and using the trained residual network to perform target recognition on the cleared spatial target image both include the following steps:

[0122] First, the sharpened spatial target image passes through a structure without residual blocks, performing convolution, regularization, activation function, and maximum pooling on the input spatial target image. Feature extraction is then performed through a four-layer structure containing residual blocks. This is then converted into a feature vector through average pooling and fully connected layers, and finally a classifier outputs the spatial target type. The "non-residual structure" refers to a structure consisting of convolutional and pooling layers, while the "residual structure" refers to a structure consisting of convolutional and pooling layers connected by residuals.

[0123] The recognition method of this embodiment first performs a preliminary positioning of the space target to find out its position, then uses the imaging prior condition (i.e., the space target image and the background environment are consistent) to de-motion blur and sharpen the located space target image, and finally uses the residual network to perform target recognition, thereby removing interference in the space target image and restoring the space target image with rich detail information. The progressive method of first positioning and then recognition improves the accuracy of space target detection and recognition in wide images.

[0124] Example 2

[0125] Based on the first embodiment, this embodiment verifies and illustrates the effect of the imaging prior-driven progressive spatial small target intelligent recognition method described in the first embodiment through simulation experiments.

[0126] Test conditions:

[0127] This example uses the BUAA-SID-Share1.0 dataset for testing. This dataset contains 20 different types of space targets, each with 197 training images and 33 test images, for a total of 4,600 grayscale images. This dataset was blurred and then added to a wide-angle deep-space background to form the test set for this example.

[0128] See Figure 2 、 Figure 3 and Figure 4 , Figure 2 A schematic diagram of a space target in the BUAA-SID-Share1.0 dataset provided by an embodiment of the present invention, Figure 3 A schematic diagram of a space target after fuzzy processing provided by an embodiment of the present invention, Figure 4 This is a schematic diagram of an image of a blurred space target added to a wide deep space background provided by an embodiment of the present invention. Figure 2 、 Figure 3 and Figure 4 The spatial target data set with motion blur to be identified is obtained through processing.

[0129] Test content:

[0130] First, a wide image of a motion-blurred spatial target is input to locate the motion-blurred spatial target. The blurred and real images are then used to train the degradation model network and the improved enhanced super-resolution generative adversarial network until convergence, achieving deblurring and sharpening of the motion-blurred spatial target. Finally, the residual network is trained until convergence to achieve classification and recognition of the spatial target.

[0131] Experimental results and analysis:

[0132] See Figure 5 、 Figure 6 and Table 1, Figure 5 It is a schematic diagram of the located space target image with motion blur. Figure 6 is a schematic diagram of a space target image after deblurring and sharpening operations, and Table 1 is a recognition result of the space target image after deblurring and sharpening operations.

[0133] Table 1 Recognition results of spatial target images after deblurring and sharpening operations

[0134]

[0135] Depend on Figure 5 It can be seen that the located space target image has motion blur, which results in less local detail information. Direct recognition will lead to lower accuracy. Figure 6 As shown in the figure, after deblurring and sharpening, the local details of the spatial target are restored, more closely matching the actual image, providing richer information for subsequent recognition. Table 1 shows that the progressive recognition method, which uses localization followed by detection and leverages prior knowledge, achieves high recognition accuracy for multiple categories of spatial targets (with the lowest accuracy reaching 97%).

[0136] This embodiment improves the accuracy of space target detection and recognition by first locating a small space target with motion blur, determining its position, then using the consistency of the space target image background environment to remove motion blur, and performing a sharpening operation at the same time, and finally using a residual network to identify the space target with rich detail information. This progressive recognition method improves the accuracy of space target detection and recognition.

[0137] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A progressive small-target intelligent recognition method driven by imaging priors, characterized by: Including steps: S1. Use the minimum bounding rectangle algorithm based on threshold embedding to preliminarily locate the spatial target with motion blur and obtain a preliminary positioning image; S2. Extracting degradation features of the preliminary positioning image and calculating the residual between the preliminary positioning image and the degradation features to obtain a deblurred spatial target image; including: randomly cropping the preliminary positioning image to obtain a first image block and a second image block; performing data enhancement on both the first image block and the second image block to obtain a first enhanced image and a second enhanced image; performing feature extraction on both the first enhanced image and the second enhanced image using an encoder to obtain a first extracted feature and a second extracted feature; inputting the first extracted feature and the second extracted feature into a fully connected layer and maximizing the mutual information between the first extracted feature and the second extracted feature to obtain the degradation feature; calculating the residual between the preliminary positioning image and the degradation feature to obtain a deblurred spatial target image; S3. Using the trained improved enhanced super-resolution generative adversarial network to perform feature extraction on the deblurred space target image to obtain a clear space target image; S4. Utilize the trained residual network to perform target recognition on the cleared space target image to obtain the space target type.

2. The imaging prior-driven progressive small target intelligent recognition method according to claim 1, characterized in that: Step S1 includes: The area where the space target is located is determined according to the outline of the space target, the maximum point and the minimum point of all pixel values ​​in the area are obtained, and a rectangular frame is made according to the maximum point and the minimum point to obtain the preliminary positioning image.

3. The imaging prior-driven progressive small target intelligent recognition method according to claim 1, characterized in that: Step S1 includes: Determine several regions where the space target is located according to the outline of the space target, obtain the maximum value point and the minimum value point of all pixel values ​​in each of the regions, and draw rectangular frames based on the maximum value point and the minimum value point to obtain several rectangular frames; The plurality of rectangular frames are merged according to their intersection-over-union ratio to obtain the preliminary positioning image, and the merging rule is: in, and Respectively represent the horizontal and vertical coordinates of the lower left corner of the first rectangular box, and They represent the horizontal and vertical coordinates of the lower left corner of the second rectangular box, IOU represents the intersection-over-union ratio of the two rectangular boxes, and threshold represents the set threshold.

4. The imaging prior-driven progressive small target intelligent recognition method according to claim 1, characterized in that: The improved enhanced super-resolution generative adversarial network includes a generator and a discriminator connected in sequence; Step S3 includes: S31. Training the generator and the discriminator using a low-resolution image and a real high-resolution image corresponding to the low-resolution image to obtain the trained improved enhanced super-resolution generative adversarial network, wherein the low-resolution image is a deblurred spatial target image obtained by sequentially undergoing preliminary positioning and deblurring processing; S32. Input the deblurred spatial target image into the trained improved enhanced super-resolution generative adversarial network, and use the trained generator and discriminator to perform feature extraction in sequence to obtain the cleared spatial target image.

5. The imaging prior-driven progressive small target intelligent recognition method according to claim 4, characterized in that: Step S31 includes: S311, inputting the low-resolution image into the convolution layer, the rectified linear unit, the multiple dense residual blocks and the upsampling network in the generator in sequence for feature extraction to obtain a reconstructed high-resolution image; S312, using the output of the discriminator in the previous iterative step and the reconstructed high-resolution image to calculate the adversarial loss of the generator, and using the adversarial loss of the generator to update the parameters of the generator to obtain an updated generator; S313, inputting the real high-resolution image and the reconstructed high-resolution image into the discriminator for feature extraction to obtain the output of the discriminator; wherein the discriminator adopts a VGG network; S314, calculating the adversarial loss of the discriminator using the output of the discriminator, the true high-resolution image, and the reconstructed high-resolution image, and training the discriminator using the adversarial loss of the discriminator to obtain an updated discriminator; S315. Calculate the perceptual loss of the improved enhanced super-resolution generative adversarial network using the low-resolution image and the reconstructed high-resolution image during training, and optimize the improved enhanced super-resolution generative adversarial network using the perceptual loss; S316. When the convergence condition is reached, the trained improved enhanced super-resolution generative adversarial network is obtained.

6. The imaging prior-driven progressive small target intelligent recognition method according to claim 5, characterized in that: The adversarial loss of the generator is: in, represents the adversarial loss function of the generator, and represent the low-resolution image and the reconstructed high-resolution image, respectively. and Respectively represent the average of image data within a batch of low-resolution images and reconstructed high-resolution images, , , represents the sigmoid function, Represents the output of the discriminator in the previous iteration step.

7. The imaging prior-driven progressive small target intelligent recognition method according to claim 5, characterized in that: The adversarial loss of the discriminator is: in, represents the loss function of the relative average discriminator, and represent the real high-resolution image and the reconstructed high-resolution image, respectively. and Respectively represent the mean of image data within a batch of real high-resolution images and reconstructed high-resolution images, , , represents the sigmoid function, Represents the output of the relative average discriminator.

8. The imaging prior-driven progressive small target intelligent recognition method according to claim 5, characterized in that: The perceptual loss includes content loss and generation loss, wherein, The content loss is Euclidean distance loss; The generation loss is: in, represents the probability that the discriminator identifies the reconstructed high-resolution image as the real high-resolution image.

9. The imaging prior-driven progressive small target intelligent recognition method according to claim 1, characterized in that: The residual network includes a ResNet-50 network.

Images