Self-adaptive image super-resolution reconstruction method of intelligent reconnaissance equipment

Abstract

The invention relates to the technical field of computer vision, and provides a self-adaptive image super-resolution reconstruction method for intelligent reconnaissance equipment, which comprises the following steps of: acquiring a low-resolution image through the intelligent reconnaissance equipment; inputting the low-resolution image into a trained degradation model for feature extraction to obtain a first degradation feature; performing no-reference image quality evaluation on the low-resolution image to obtain a degradation type and a degradation degree of the image, and mapping the degradation type and the degradation degree into a vector to obtain a second degradation feature; performing feature fusion on the first degradation feature and the second degradation feature to obtain a comprehensive degradation vector feature; inputting the low-resolution image and the comprehensive degradation vector feature into a super-resolution reconstruction model based on depth separable convolution for processing to obtain a reconstructed high-resolution image; wherein the super-resolution reconstruction model performs adaptive adjustment of the number of network layers according to the comprehensive degradation vector features.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and more specifically, to an adaptive image super-resolution reconstruction method for intelligent reconnaissance equipment. Background Art

[0002] As an observation device installed on carriers such as drones, intelligent reconnaissance equipment can carry a variety of different types of sensors and has demonstrated its unique application value in many fields. Whether in military reconnaissance, disaster monitoring, environmental assessment, or urban planning, intelligent reconnaissance equipment can provide vital image and data support. The accuracy and clarity of these image data directly affect the subsequent analysis and application effects. However, due to various factors such as flight speed, camera shake, atmospheric interference, etc., the images obtained by intelligent reconnaissance equipment often have problems such as low resolution and blurred details. Among them, image super-resolution technology can restore high-resolution images from low-resolution images, thereby improving the clarity and availability of images. Therefore, image super-resolution technology has a very important application prospect in the field of intelligent reconnaissance equipment.

[0003] Due to the complexity of flight conditions, images acquired by intelligent reconnaissance equipment may contain multiple types of degradation, such as motion blur, noise, illumination changes, etc. These complex degradation conditions make it difficult for traditional super-resolution methods to effectively restore image details. Secondly, existing super-resolution methods usually rely on predefined degradation models, which may not be able to adapt to the complex degradation conditions of intelligent reconnaissance equipment images. In addition, the application scenarios of intelligent reconnaissance equipment usually require real-time processing of large amounts of image data, which places higher requirements on the efficiency and computing resources of super-resolution algorithms. Therefore, how to effectively solve these problems and improve the performance and efficiency of super-resolution reconstruction of images of intelligent reconnaissance equipment has important research significance and application value. Summary of the invention

[0004] In order to overcome the defect that the above-mentioned existing super-resolution reconstruction method is difficult to effectively restore image details, the present invention provides an adaptive image super-resolution reconstruction method for intelligent reconnaissance equipment.

[0005] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0006] An adaptive image super-resolution reconstruction method for intelligent reconnaissance equipment comprises the following steps:

[0007] Collect low-resolution images through intelligent reconnaissance equipment;

[0008] Inputting the low-resolution image into a trained degradation model to extract features and obtain a first degradation feature;

[0009] Performing a no-reference image quality assessment on the low-resolution image, obtaining the degradation type and degradation degree of the image, and then mapping the obtained degradation type and degradation degree into a vector to obtain a second degradation feature;

[0010] Fusing the first degradation feature and the second degradation feature to obtain a comprehensive degradation vector feature;

[0011] The low-resolution image and the comprehensive degradation vector feature are input into a super-resolution reconstruction model based on depthwise separable convolution for processing to obtain a reconstructed high-resolution image; wherein the super-resolution reconstruction model adaptively adjusts the number of network layers according to the comprehensive degradation vector feature.

[0012] Furthermore, the present invention also proposes a device, including intelligent reconnaissance equipment, wherein a memory and a processor are deployed on the intelligent reconnaissance equipment, and computer-readable instructions are stored in the memory. When the computer-readable instructions are executed by the processor, the processor executes all or part of the steps of the adaptive image super-resolution reconstruction method proposed in the present invention.

[0013] Furthermore, the present invention also proposes a storage medium having computer-readable instructions stored thereon, wherein the computer-readable instructions, when executed by a processor, implement all or part of the steps of the adaptive image super-resolution reconstruction method proposed in the present invention.

[0014] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0015] The present invention fuses the degradation features extracted by the degradation model and the quality assessment results of the reference-free image to obtain a comprehensive degradation vector feature, which is further used for super-resolution reconstruction, and can more comprehensively understand and compensate for image degradation, improve the accuracy and depth of feature extraction, and thus enhance the quality and reliability of the reconstructed image;

[0016] The super-resolution reconstruction model in the present invention adaptively adjusts the number of network layers according to the comprehensive degradation vector characteristics, that is, the network structure and parameters can be optimized in real time according to different degradation vectors to adapt to the current image degradation situation, thereby improving the generalization ability of the super-resolution reconstruction model, so that intelligent reconnaissance equipment can stably provide high-quality image output in changing real-world environments, such as urban monitoring, border patrols, disaster relief and other scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The figure is a flow chart of an adaptive image super-resolution reconstruction method according to an embodiment of the present invention.

[0018] Figure 2The figure is a schematic flow chart of an adaptive image super-resolution reconstruction method according to an embodiment of the present invention.

[0019] Figure 3 The figure is a training flow chart according to an embodiment of the present invention. DETAILED DESCRIPTION

[0020] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0021] The terms used in the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0022] It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0023] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] Example 1

[0025] This embodiment proposes an adaptive image super-resolution reconstruction method for intelligent reconnaissance equipment, such as Figure 1 , 2 , which is a flow chart of the adaptive image super-resolution reconstruction method of this embodiment.

[0026] The adaptive image super-resolution reconstruction method proposed in this embodiment includes the following steps:

[0027] S1. Collect low-resolution images through intelligent reconnaissance equipment;

[0028] S2, inputting the low-resolution image into a trained degradation model to extract features, thereby obtaining a first degradation feature;

[0029] S3, performing a non-reference image quality assessment on the low-resolution image, obtaining the degradation type and degradation degree of the image, and then mapping the obtained degradation type and degradation degree into a vector to obtain a second degradation feature;

[0030] S4, fusing the first degradation feature and the second degradation feature to obtain a comprehensive degradation vector feature;

[0031] S5. Input the low-resolution image and the comprehensive degradation vector feature into a super-resolution reconstruction model based on deep separable convolution for processing to obtain a reconstructed high-resolution image; wherein the super-resolution reconstruction model adaptively adjusts the number of network layers according to the comprehensive degradation vector feature.

[0032] This embodiment uses the degradation degree of the low-resolution image obtained by no-reference image quality assessment as an explicit feature, and combines it with the degradation representation learned by the degradation model as an implicit feature, which is further used for super-resolution reconstruction. It can more comprehensively understand and compensate for image degradation, effectively improve the accuracy and depth of feature extraction, and thus enhance the quality and reliability of the reconstructed image.

[0033] The super-resolution reconstruction model in this embodiment adaptively adjusts the number of network layers according to the comprehensive degradation vector characteristics, that is, the network structure and parameters can be optimized in real time according to different degradation vectors to adapt to the current image degradation situation, thereby improving the generalization ability of the super-resolution reconstruction model, so that intelligent reconnaissance equipment can stably provide high-quality image output in changing real-world environments, such as urban monitoring, border patrols, disaster relief and other scenarios.

[0034] In an optional embodiment, the intelligent reconnaissance equipment is equipped with a visible light image sensor and an edge computing device, wherein the visible light image sensor is used to collect low-resolution images, and then the collected images are transmitted to the edge computing device through a high-speed data transmission interface for image super-resolution reconstruction.

[0035] For example, the intelligent reconnaissance equipment is equipped with an sCMOS camera with an ultra-high resolution of 3200×3200, which can capture images at a speed of 498 frames per second while maintaining a quantum efficiency of up to 95%. The design of this camera meets the strict requirements for image details and acquisition speed, and can obtain high-quality images under low-light conditions. It is particularly suitable for the needs of intelligent reconnaissance equipment for high-definition images in fast and dynamic environments. In addition, the high-speed acquisition capability enables the camera to continuously record fast-occurring events without motion blur or loss of key details, ensuring that image data with rich details and extremely high clarity can be obtained, providing ideal input material for subsequent image super-resolution reconstruction.

[0036] For example, the intelligent reconnaissance equipment is equipped with an edge computing device, Jetson AGX OrinDeveloper Kit (64G), which has powerful hardware performance and flexible interface design, and can process high-resolution image data in real time. Through this efficient image acquisition and processing process, the intelligent reconnaissance equipment can provide high-quality original image data for image super-resolution reconstruction tasks, thereby significantly improving the clarity and detail of the reconstructed image and meeting the real-time needs of the intelligent reconnaissance equipment when performing tasks.

[0037] In an optional embodiment, the training process of the super-resolution reconstruction model includes the following steps:

[0038] Step 1: Collect high-resolution image samples, perform random degradation operations on the high-resolution image samples to obtain low-resolution image samples, and construct HR-LR image pairs as training data;

[0039] Step 2: Input the low-resolution image samples obtained through random degradation operation into the degradation model for feature extraction to obtain implicit degradation features;

[0040] Step 3: taking the degradation operation records and degradation parameters of the low-resolution image samples obtained through the random degradation operation as explicit degradation features;

[0041] Step 4: Fusing the implicit degradation features and the explicit degradation features to obtain a comprehensive degradation vector feature, and inputting the feature and the low-resolution image sample into the super-resolution reconstruction model for training to obtain a trained super-resolution reconstruction model.

[0042] like Figure 3 As shown, it is the degradation model training flow chart of this embodiment.

[0043] Among them, the explicit degradation feature model can directly learn the impact of the degradation operation, while the implicit degradation feature can help the model understand the interaction and overall impact between the degradation operations.

[0044] The combination of dual degradation features selected in this embodiment can enhance the adaptability of the model to unknown or complex degradation situations, improve its robustness and generalization ability in practical applications, and further enable intelligent reconnaissance equipment to more comprehensively understand and compensate for image degradation during super-resolution reconstruction, thereby improving the accuracy and depth of feature extraction, and thereby enhancing the quality and reliability of the reconstructed image.

[0045] Furthermore, in the test phase, a non-reference image quality assessment is performed on the low-resolution image samples used as test data, and the image degradation degree is evaluated by calculating a series of objective indicators to obtain the explicit degradation of the low-resolution image, such as blur, noise, compression, etc. These degradation types and degradation degrees are then mapped into vectors as explicit degradation features, which are further combined with the implicit degradation features extracted by the degradation model to obtain comprehensive degradation vector features, which are then input into the super-resolution reconstruction model to complete the model test.

[0046] In this embodiment, in order to simulate the degradation process in the real world, a series of degradation operations and their parameters are predefined to degrade the high-resolution image into a low-resolution image. This can more realistically simulate the degradation conditions in different scenarios, so that the trained model can better adapt to inputs of various degradation types.

[0047] Further, in an optional embodiment, performing a random degradation operation on the high-resolution image sample in step 1 to obtain a low-resolution image sample includes the following steps:

[0048] Step 1-1: For any high-resolution image I HR , randomly select K degradation operations from the preset degradation operations, and randomly select degradation parameters from the value range [0,1] for each degradation operation, perform the corresponding degradation operation, and obtain K degraded low-resolution image samples I LR,k , k = 1, 2, ..., K, where K is the preset number of degradation operations;

[0049] Step 1-2: Record each low-resolution image sample I LR,k The corresponding degradation operation and its parameters are used as the explicit degradation feature Among them, O k Represents a low-resolution image sample I LR,k The selected degradation operation, Denotes the degenerate operation O k The corresponding degradation parameter.

[0050] As an exemplary illustration, in step 1, a series of degradation operations are predefined, such as blur, noise, blur, etc., which are represented as O1, O2, O3, ..., O M , where each degradation operation has a corresponding degradation parameter P1, P2, P3, ..., P M , which is used to control the degree of the degradation operation, and the value range of each parameter is 0-1, and its specific value is random. The expression is as follows:

[0051]

[0052] Among them, Random(a,b) means generating a random number in the interval [a,b]. Represents the degradation operation O k Randomly selected degradation parameters. HR Randomly select the degradation operation and its parameters to generate the corresponding low-resolution image I LR , the specific formula is:

[0053]

[0054] Among them, DegradationOperations(·) represents a series of degradation operations on the image.

[0055] In this embodiment, the degradation model predefines various degradation conditions that may be encountered, such as blur, compression, and noise, so that in the super-resolution reconstruction process, the degradation effects in the low-resolution image can be more realistically simulated and compensated, thereby improving the quality and detail expression of the reconstructed image. At the same time, the image processing capability of the intelligent reconnaissance equipment in complex environments is enhanced, and more accurate visual information is provided for subsequent image analysis and decision-making.

[0056] Furthermore, in an optional embodiment, in step 2, the low-resolution image samples obtained by the random degradation operation are input into the degradation model for feature extraction; the steps include:

[0057] Step 2-1, cropping and dividing the input low-resolution image samples;

[0058] Step 2-2: input the image blocks into the degradation model to extract features, and obtain degradation features of each image block.

[0059] Traditional super-resolution methods often rely on accurate estimation of degradation models, which requires a lot of prior knowledge and manual adjustments. However, in practical applications, the degradation process of an image may be very complex and difficult to estimate accurately. By extracting and fusing known degradation information, this embodiment allows the model to learn abstract features of the degradation process, providing a more flexible and adaptive method for handling unknown or complex degradation situations.

[0060] As an exemplary explanation, this embodiment assumes that the degradation type in the same image is the same, and the degradation of different images is different. In the feature extraction process, the low-resolution image is first segmented into image blocks and then feature extraction is performed. For each input low-resolution image, it is divided into 196x 196 image blocks, and feature extraction is performed on each image block.

[0061] Exemplarily, the degradation model includes a ResNet50 network. Then its expression is:

[0062] F i =ResNet50(X i )

[0063] Among them, X i Represents a low-resolution image I LR The i-th image block in F i Represents the features extracted from the i-th image block.

[0064] Furthermore, in an optional embodiment, the method further comprises the following steps:

[0065] An unsupervised learning method is used to learn the degradation feature representation of the degradation model, and a contrastive learning method is used to distinguish different degradation representations in the feature space; wherein different image blocks X from the same image are i and X j As positive sample pairs, different image patches X from different images are i and X k As negative sample pairs, the objective function L is constructed with the goal of maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs. contrastive ; Its expression is:

[0066]

[0067] Among them, F i ,F j ,F k Represents image blocks X i ,X j ,X k The degradation characteristics of ; τ is the temperature parameter used to control the contrast between positive and negative samples; sim(·,·) is the similarity calculation function.

[0068] In this embodiment, it is assumed that the degradation types of different image blocks corresponding to the same image are the same.

[0069] This embodiment uses a contrastive learning method to distinguish different degraded representations in the feature space, with the goal of bringing different image blocks of the same image (positive sample pairs) closer together in the feature space and moving image blocks of different images (negative sample pairs) farther apart in the feature space.

[0070] Specifically, for each image block X i , the corresponding degenerate expression is F i ; Take another image block X of the image j , the corresponding degenerate expression is F j ; For another image block X k , the corresponding degenerate expression is Fk Then, relative to F i In terms of i and F j are positive samples for each other; relative to F i , F k is a negative sample. The goal of the above formula is to maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs, so that the feature vectors of the same image block are closer, while the feature vectors of different image blocks are farther away. This method can help the model better learn the difference between features, thereby improving its ability to recognize different images and degradation types.

[0071] Further, in an optional embodiment, performing no-reference image quality assessment on the low-resolution image sample in step S3 includes the following steps:

[0072] Calculate the noise level of the image sample; its expression is:

[0073]

[0074] Among them, noise_score represents the noise level score of the image, std(I LR ) represents image I LR The standard deviation, mean(I LR ) represents image I LR The average value of

[0075] Calculate the compression level of image samples; the expression is:

[0076]

[0077] Among them, compression_ratio represents the compression degree score of the image, Originalsize(I LR ) represents the size of the original image, Compressedsize(I LR ) represents the compressed image size;

[0078] Calculate the blur noise of image samples; its expression is:

[0079]

[0080] Among them, blue_score represents the blur score of the image, gradient_magnitude(I LR ) represents the gradient size of the image, H represents the height of the image, and W represents the width of the image; the gradient size can be obtained by calculating the gradient of the low-resolution image in the horizontal and vertical directions; its expression is:

[0081]

[0082] Among them, G x,LR and G y,LR Respectively represent the gradient of the image in the horizontal and vertical directions, I LR,i,j represents the pixel value of the low image at position (i, j), H and W represent the height and width of the image respectively;

[0083] The noise level score, compression level score and blur level score of the image are mapped into vectors to obtain explicit degradation vector features.

[0084] Further, in an optional embodiment, the super-resolution reconstruction model in step S5 performs adaptive adjustment of the number of network layers according to the comprehensive degradation vector feature, comprising the following steps:

[0085] The number of network layers L is increased or decreased according to the comprehensive degradation vector feature, and the super-resolution reconstruction model is adjusted; the expression is:

[0086]

[0087] Among them, L0 is the initial layer number of the basic network, Denotes the degenerate operation O k The corresponding degradation parameter; f layers (·) is used to dynamically adjust the number of network layers according to the degree of different degradation types; τ1, τ2 are preset degradation thresholds; λ1, λ2 are incremental coefficients used to determine the amount of increase in the number of network layers;

[0088] The number of filters F of each layer of the basic network is determined according to the comprehensive degradation vector feature, and the super-resolution reconstruction model is adjusted to enhance the learning ability of the feature degradation feature; its expression is:

[0089]

[0090] Among them, F0 is the initial number of filters in each layer of the basic network, Denotes the degenerate operation O k The corresponding degradation parameter; g filters (·) is used to dynamically adjust the number of filters according to the degree of different degradation types; λ3,λ4 are incremental coefficients, which are used to determine the increase in the number of filters;

[0091] The number of jump connections SC in the basic network is determined according to the comprehensive degradation vector feature, and the super-resolution reconstruction model is adjusted; its expression is:

[0092]

[0093] Where, SC0 is the number of initial jump connections of the basic network, T is the preset degradation threshold; D full is the overall degradation degree of the image; h sc (·) is used to dynamically adjust the number of skip connections according to the overall degradation of the image; α is the rate at which the skip connections grow, and is used to adjust the amount of skip connections added. If the degradation vector indicates that the image is severely degraded, the information flow from deep layers to shallow layers can be strengthened by adding skip connections; for slightly degraded images, the number of skip connections can be reduced to simplify the network structure.

[0094] In this embodiment, a piecewise function is selected to dynamically adjust the number of network layers and filters according to the uneven degree of degradation. Specifically, by setting degradation thresholds τ1, τ2, the network will be adjusted according to different intervals of degradation characteristics. When it is less than the threshold τ1, the network does not increase the number of additional layers and filters; when When it is in the interval [τ1,τ2], the network increases the number of layers and the number of filters; When the threshold τ2 is exceeded, the network will significantly increase the number of layers and the number of filters.

[0095] Among them, the increment of the number of layers and filters in each interval can be adjusted by setting λ1,λ2,λ3,λ4 to cope with different degradation complexities, thereby enhancing the adaptability and generalization ability of the network while maintaining computational efficiency.

[0096] Furthermore, according to the relationship between the overall degradation degree of the image and the preset threshold, when D full >T, it indicates that the image is severely degraded and the information is greatly lost. It is necessary to increase the number of skip connections to further restore the image details and improve the model's generalization and recovery capabilities for images. Otherwise, the current number of skip connections remains unchanged, indicating that the current image degradation is within an acceptable range.

[0097] This embodiment dynamically adjusts the network structure according to the degradation vector, so that the model can better cope with various complex degradation situations and improve the effect and quality of super-resolution reconstruction of low-resolution images. This adaptive adjustment mechanism significantly improves the generalization ability of image processing of intelligent reconnaissance equipment in various complex environments, ensures the high-quality output of super-resolution reconstructed images, and can provide clear visual effects and accurate detail information in both long-distance reconnaissance and monitoring tasks under low-light conditions.

[0098] Further, in an optional embodiment, the super-resolution reconstruction model in step S5 includes the following steps when performing super-resolution reconstruction:

[0099] Perform a deep separable convolution operation on the input comprehensive degradation vector feature to obtain a high-resolution feature F HR ;

[0100] For the high-resolution feature F HR Upsample to generate high-resolution images Its expression is:

[0101]

[0102] Where M′ is a set of modulation coefficients used to modulate the features of different channels; Upsample(·) represents the upsampling operation.

[0103] Optionally, a loss function is designed to measure the difference between the reconstructed image and the true high-resolution image. The specific formula is as follows:

[0104]

[0105] Where λ is the regularization term weight and Ω(P) is the regularization term of the network parameters, including the modulation coefficient M′.

[0106] In the super-resolution reconstruction process of this embodiment, the intelligent reconnaissance equipment can not only handle common types of image degradation, but also adapt to unknown or multiple composite degradation situations. This improvement in generalization capability enables the intelligent reconnaissance equipment to stably provide high-quality image output in changing real-world environments, such as urban monitoring, border patrols, disaster relief and other scenarios. In addition, this generalization capability also means that the intelligent reconnaissance equipment can quickly adapt to new degradation patterns without a large amount of retraining data, thereby improving the practicality and response speed of the intelligent reconnaissance equipment.

[0107] Example 2

[0108] This embodiment proposes a device, including intelligent reconnaissance equipment, on which a memory and a processor are deployed, and the memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor executes all or part of the steps of the adaptive image super-resolution reconstruction method proposed in Example 1.

[0109] It can be understood that the system of this embodiment corresponds to the method of the above-mentioned embodiment 1, and the options in the above-mentioned embodiment 1 are also applicable to this embodiment, so they will not be described repeatedly here.

[0110] Example 3

[0111] This embodiment proposes a storage medium on which computer-readable instructions are stored, wherein the computer-readable instructions, when executed by a processor, implement all or part of the steps of the adaptive image super-resolution reconstruction method proposed in Embodiment 1.

[0112] Exemplarily, the storage medium includes, but is not limited to, a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0113] Exemplarily, the instructions, programs, code sets or instruction sets may be implemented using conventional programming languages.

[0114] Exemplarily, the processor includes but is not limited to a smart phone, a personal computer, a server, a network device, etc., and is used to execute all or part of the steps of the adaptive image super-resolution reconstruction method described in Example 1.

[0115] The terms in the drawings are only used for illustrative purposes and are not to be construed as limiting the present invention;

[0116] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. An adaptive image super-resolution reconstruction method for intelligent reconnaissance equipment, characterized in that: The following steps are involved: Collect low-resolution images through intelligent reconnaissance equipment; Inputting the low-resolution image into a trained degradation model to extract features and obtain a first degradation feature; Performing a no-reference image quality assessment on the low-resolution image, obtaining the degradation type and degradation degree of the image, and then mapping the obtained degradation type and degradation degree into a vector to obtain a second degradation feature; Fusing the first degradation feature and the second degradation feature to obtain a comprehensive degradation vector feature; The low-resolution image and the comprehensive degradation vector feature are input into a super-resolution reconstruction model based on depthwise separable convolution for processing to obtain a reconstructed high-resolution image; wherein the super-resolution reconstruction model adaptively adjusts the number of network layers according to the comprehensive degradation vector feature.

2. The adaptive image super-resolution reconstruction method according to claim 1, characterized in that: The training process of the super-resolution reconstruction model includes the following steps: Collect high-resolution image samples, perform random degradation operations on the high-resolution image samples to obtain low-resolution image samples, and construct HR-LR image pairs as training data; The low-resolution image samples obtained through random degradation operation are input into the degradation model for feature extraction to obtain implicit degradation features; The degradation operation records and degradation parameters of the low-resolution image samples obtained by random degradation operation are used as explicit degradation features; The implicit degradation features and the explicit degradation features are feature fused to obtain a comprehensive degradation vector feature, which is input into the super-resolution reconstruction model with the low-resolution image sample for training to obtain a trained super-resolution reconstruction model.

3. The adaptive image super-resolution reconstruction method according to claim 2, characterized in that: The step of performing a random degradation operation on the high-resolution image sample to obtain a low-resolution image sample comprises the following steps: For any high-resolution image I HR , randomly select K degradation operations from the preset degradation operations, and randomly select degradation parameters from the value range [0,1] for each degradation operation, perform the corresponding degradation operation, and obtain K degraded low-resolution image samples I LR,k , k=1,2,...,K; Record each low-resolution image sample I LR,k The corresponding degradation operation and its parameters are used as the explicit degradation feature Among them, O k Represents a low-resolution image sample I LR,k The selected degradation operation, Denotes the degenerate operation O k The corresponding degradation parameter.

4. The adaptive image super-resolution reconstruction method according to claim 2, characterized in that: The step of inputting the low-resolution image samples obtained through random degradation operation into the degradation model for feature extraction comprises the following steps: Crop and divide the input low-resolution image samples into blocks; The image blocks are input into the degradation model to extract features, and the degradation features of each image block are obtained.

5. The adaptive image super-resolution reconstruction method according to claim 4, characterized in that: The method further comprises the following steps: An unsupervised learning method is used to learn the degradation feature representation of the degradation model, and a contrastive learning method is used to distinguish different degradation representations in the feature space; wherein different image blocks X from the same image are i and X j As positive sample pairs, different image patches X from different images are i and X k As negative sample pairs, the objective function L is constructed with the goal of maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs. contrastive ; Its expression is: Among them, F i ,F j ,F k Represents image blocks X i ,X j ,X k The degradation characteristics of ; τ is the temperature parameter used to control the contrast between positive and negative samples; sim(·,·) is the similarity calculation function.

6. The adaptive image super-resolution reconstruction method according to claim 1, characterized in that: The step of performing no-reference image quality assessment on the low-resolution image comprises the following steps: Calculate the noise level of the image sample; its expression is: Among them, noise_score represents the noise level score of the image, std(I LR ) represents image I LR The standard deviation, mean(I LR ) represents image I LR The average value of Calculate the compression level of image samples; the expression is: Among them, compression_ratio represents the compression degree score of the image, Originalsize(I LR ) represents the size of the original image, Compressedsize(I LR ) represents the compressed image size; Calculate the blur noise of the image sample; its expression is: Among them, blue_score represents the blur score of the image, gradient_magnitude(I LR ) represents the gradient size of the image, H represents the height of the image, and W represents the width of the image; The noise level score, compression level score and blur level score of the image are mapped into vectors to obtain explicit degradation vector features.

7. The adaptive image super-resolution reconstruction method according to claim 1, characterized in that: The super-resolution reconstruction model adaptively adjusts the number of network layers according to the comprehensive degradation vector feature, comprising the following steps: The number of network layers L is increased or decreased according to the comprehensive degradation vector feature, and the super-resolution reconstruction model is adjusted; the expression is: Among them, L0 is the initial layer number of the basic network, Denotes the degenerate operation O k The corresponding degradation parameter; f layers (·) is used to dynamically adjust the number of network layers according to the degree of different degradation types; τ1, τ2 are preset degradation thresholds; λ1, λ2 are incremental coefficients used to determine the increase in the number of network layers; The number of filters F of each layer of the basic network is determined according to the comprehensive degradation vector feature, and the super-resolution reconstruction model is adjusted; its expression is: Among them, F0 is the initial number of filters in each layer of the basic network, Denotes the degenerate operation O k The corresponding degradation parameter; g filters (·) is used to dynamically adjust the number of filters according to the degree of different degradation types; λ3,λ4 are incremental coefficients, which are used to determine the increase in the number of filters; The number of jump connections SC in the basic network is determined according to the comprehensive degradation vector feature, and the super-resolution reconstruction model is adjusted; its expression is: Among them, SC0 is the initial jump connection number of the basic network, T is the preset degradation threshold; D full is the overall degradation degree of the image; h sc (·) is used to dynamically adjust the number of skip connections according to the overall degradation of the image; α is the rate at which the skip connections grow, which is used to adjust the amount of skip connections added.

8. The adaptive image super-resolution reconstruction method according to any one of claims 1 to 7, characterized in that: The super-resolution reconstruction model comprises the following steps when performing super-resolution reconstruction: Perform a deep separable convolution operation on the input comprehensive degradation vector feature to obtain a high-resolution feature F HR ; For the high-resolution feature F HR Upsample to generate high-resolution images Its expression is: Where M′ is a set of modulation coefficients used to modulate the features of different channels; Upsample(·) represents the upsampling operation.

9. A device, comprising intelligent reconnaissance equipment, characterized in that: The intelligent reconnaissance equipment is deployed with a memory and a processor, the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes all or part of the steps of the adaptive image super-resolution reconstruction method according to any one of claims 1 to 8.

10. A storage medium having computer-readable instructions stored thereon, characterized in that: When the computer-readable instructions are executed by a processor, all or part of the steps of the adaptive image super-resolution reconstruction method according to any one of claims 1 to 8 are implemented.

Images