An anti-distillation image restoration method and system based on adaptive singular value perturbation

Abstract

This invention discloses a method and system for backdistillation image restoration based on adaptive singular value perturbation, relating to the field of computer vision recognition technology. The method includes: training a backdistillation model composed of a teacher model and a student model using training images; the training process includes: extracting multi-scale features from the training images in parallel using both the teacher and student models; performing singular value decomposition on the intermediate layer features generated by the teacher model, amplifying the first principal singular values ​​while leaving the remaining singular values ​​unchanged, thereby reconstructing a perturbation feature map; inputting the perturbation feature map into the corresponding layer of the student model for feature distillation; and applying the trained backdistillation model to the image to be restored to obtain an enhanced image. Through structured singular value perturbation and dual-path separation design, the method achieves strong anti-distillation and lossless teacher model quality.

Description

Technical Field

[0001] This invention relates to the field of computer vision recognition technology, and in particular to a method and system for back-distillation image restoration based on adaptive singular value perturbation. Background Technology

[0002] Currently, knowledge distillation (KD) is widely used in image restoration to achieve model compression and acceleration. Its principle is to use the output of a high-performance teacher model as a soft label to guide a lightweight student model in learning to reconstruct mappings, thereby achieving high-quality image restoration results in resource-constrained scenarios such as mobile devices. However, this process faces security challenges: attackers only need to obtain the inference interface of the teacher model to replicate its key capabilities through distillation training, thus stealing the model or its performance. This black-box distillation attack has become one of the most significant security vulnerabilities facing image restoration.

[0003] Currently, backdistillation methods are mainly applied to image classification, typically misleading student models by injecting noise, sparsification, or adversarial perturbations into the probability distribution of the teacher's output. However, image restoration tasks differ fundamentally from classification tasks in output format and learning mechanisms. Image restoration tasks output continuous, high-dimensional images, requiring the preservation of pixel-level details and spatial consistency. Therefore, slight perturbations cannot prevent student models from learning true restoration mappings, while excessive perturbations directly damage the image quality of the teacher model. For example, in low-light enhancement tasks, when high-intensity noise or channel dropping is added to the teacher features, the visual quality of the teacher model significantly deteriorates, but the student model can still learn relatively stable restoration features. Therefore, output perturbation strategies for image classification tasks cannot be transferred to image restoration tasks.

[0004] To address these shortcomings, existing research has introduced perturbations into the intermediate feature space of the teacher model, such as adding random noise, randomly discarding feature channels, or using adversarial perturbations. However, when the perturbation intensity is too low, the student model can still effectively align with the feature representation of the teacher model; when the perturbation intensity is too high, it destroys the structural information and details of the teacher model itself, and may even lead to a significant degradation in the teacher model's output. Related experiments have shown that, regardless of whether noise is injected or channels are discarded, once the perturbation intensity reaches the level that blocks distillation, the performance of the teacher model deteriorates significantly, making it difficult to maintain the ability to reconstruct the original image. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a back-distillation image restoration method and system based on adaptive singular value perturbation. Through structured singular value perturbation and dual-path separation design, it achieves strong anti-distillation while maintaining the quality of the teacher model.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a back-distillation image restoration method based on adaptive singular value perturbation, comprising: Training images were used to train the back distillation model, which consisted of a teacher model and a student model. The training process includes: extracting multi-scale features from the training images in parallel using both the teacher model and the student model; performing singular value decomposition on the intermediate layer features generated by the teacher model; and further processing the preceding features. The principal singular values ​​are amplified while the other singular values ​​remain unchanged, and the perturbation feature map is reconstructed in this way. The perturbation feature map is then input into the corresponding layer of the student model for feature distillation. For the image to be restored, the trained back-distillation model is used to obtain the enhanced image.

[0007] As an alternative implementation, the first... Each principal singular value is amplified in a controlled manner, replacing the original singular values ​​with: , This is the magnification factor. For the first One original singular value, For the first An amplified singular value.

[0008] As an alternative implementation method, the amplification factor is adaptively determined based on the characteristic spectrum energy distribution. Specifically: before calculation The cumulative energy percentage of each principal singularity Calculate the amplification factor based on the cumulative energy percentage. in, This is the adjustment coefficient.

[0009] As an alternative implementation method, the cumulative energy percentage is: ;in, Let t be the t-th singular value, and r be the total number of singular values.

[0010] As an alternative implementation, after singular value amplification is completed, the obtained singular value amplification matrix is ​​used. and the original direction matrix and Reconstructing the perturbation feature map .

[0011] As an alternative implementation, the unperturbed original intermediate layer features are directly passed to the next network layer of the teacher model.

[0012] Secondly, the present invention provides a back-distillation image restoration system based on adaptive singular value perturbation, comprising: The training module is configured to train the backdistillation model, which consists of a teacher model and a student model, using training images. The training process includes: extracting multi-scale features from the training images in parallel using both the teacher model and the student model; performing singular value decomposition on the intermediate layer features generated by the teacher model; and further processing the preceding features. The principal singular values ​​are amplified while the other singular values ​​remain unchanged, and the perturbation feature map is reconstructed in this way. The perturbation feature map is then input into the corresponding layer of the student model for feature distillation. The recovery module is configured to use a trained back-distillation model to obtain an enhanced image from the image to be recovered.

[0013] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0014] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0015] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a backdistillation image restoration method and system based on adaptive singular value perturbation. It applies structured perturbation to the intermediate features of the teacher model based on singular value decomposition, without directly destroying the spatial structure of the features. Instead, it controllably amplifies only the top k most important principal singular values ​​of the features, thereby generating high-frequency but controllable perturbations. Simultaneously, this invention designs a dual-path structure: the amplified perturbed features are only provided to the student model as distillation supervision, while the teacher model continues inference along the unperturbed clean feature path. Therefore, it can prevent the student model's learning ability from being compromised while avoiding impacting the teacher model's output quality, achieving strong anti-distillation and lossless teacher model quality. This addresses the fundamental contradiction in current backdistillation techniques for image restoration tasks: perturbations are insufficient to prevent distillation, while perturbations damage teacher performance. Through structured singular value perturbation and a dual-path separation design, it achieves a low-cost, highly versatile, and highly stable model protection scheme.

[0017] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the principle architecture of the back-distillation image restoration method based on adaptive singular value perturbation provided in Embodiment 1 of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. Furthermore, it should be understood that the terms “comprising” and “including”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0024] Example 1 This embodiment provides a back-distillation image restoration method based on adaptive singular value perturbation, including: Training images were used to train the back distillation model, which consisted of a teacher model and a student model. The training process includes: extracting multi-scale features from the training images in parallel using both the teacher model and the student model; performing singular value decomposition on the intermediate layer features generated by the teacher model; and further processing the preceding features. The principal singular values ​​are amplified while the other singular values ​​remain unchanged, and the perturbation feature map is reconstructed in this way. The perturbation feature map is then input into the corresponding layer of the student model for feature distillation. For the image to be restored, the trained back-distillation model is used to obtain the enhanced image.

[0025] In this embodiment, the dedistillation model includes two parallel paths: a teacher model and a student model. Both start from the same input image, but the intermediate layer features of the teacher model must pass through an Adaptive Singular Value Perturbation (ASVP) module before being accessed by the student model. By inserting this module into the feature flow path, a structured perturbation of the teacher's features is achieved, thereby preventing the student model from directly imitating the teacher's true feature distribution.

[0026] The core idea of ​​this scheme is to perturb the features provided to the student model without changing the original reasoning process of the teacher model, so as to prevent unauthorized knowledge distillation.

[0027] The following is combined with Figure 1 The method of this embodiment will be described in detail. Figure 1 In the diagram, (a) represents the anti-distillation architecture, and (b) represents the singular value perturbation module.

[0028] S1: Multi-scale features are extracted from the training images in parallel using both the teacher model and the student model.

[0029] S2: Extract the feature maps of each intermediate layer generated by the teacher model. The input is sent to the ASVP module, where the intermediate layer feature maps are processed. Perform Singular Value Decomposition (SVD) to transform the intermediate layer feature maps Represented as: ; in, and All are direction matrices, describing the directional structure of the feature space; It is a diagonal matrix of singular values, containing singular values ​​ordered by importance. Its diagonal elements are singular values ​​arranged in order, i.e. It reflects the energy distribution of the feature in each major direction. The larger the singular value, the greater its contribution to the overall feature energy.

[0030] Through this decomposition process, the most representative components in the feature map are explicitly extracted, providing a basis for subsequent targeted perturbations.

[0031] S3: To prevent the student model from stably learning the feature maps, the first few elements of the singular valued diagonal matrix are... Each principal singular value is amplified in a controlled manner, that is, the original singular value is replaced with: , This is the magnification factor. For the first One original singular value, For the first One singular value is amplified; while the remaining singular values ​​remain unchanged. This operation is as follows: Figure 1 The yellow and red areas are shown in the center.

[0032] In this embodiment, the amplification factor is adaptively determined based on the energy distribution of the characteristic spectrum. , specifically: First, before calculation The cumulative energy percentage of each principal singularity ; ; in, Let t be the t-th singular value, and r be the total number of singular values.

[0033] Then, calculate the amplification factor based on the cumulative energy percentage: ; in, This is the adjustment coefficient.

[0034] current The lower the cumulative energy percentage of the principal singularities, the greater the amplification factor. The larger it is, the smaller it is.

[0035] This selective augmentation serves two important purposes: firstly, by amplifying high-energy directions, it significantly alters the learnability of features, making it difficult for student models to obtain stable gradients through backpropagation; secondly, it modifies only singular values ​​without changing the orientation matrix. , Therefore, the disturbance is structural and will not introduce instability effects caused by random noise.

[0036] S4: After completing the singular value amplification, use the original direction matrix. and and singular value magnification matrix Reconstruct the perturbation feature map: The perturbation feature map is the visible feature version of the student model.

[0037] At the same time, the original feature map that has not been disturbed The data is directly passed to the next network layer of the teacher model, ensuring that the inference process and output quality of the teacher model are completely unaffected. This creates two isolated data streams: a clean path (specific to the teacher model) and a perturbation path (specific to the student model), which can effectively limit the learning ability of the student model while maintaining the performance of the teacher model.

[0038] S5: Input the perturbation feature map into the corresponding layer of the student model for feature distillation to complete the training of the back-distillation model; then, use the trained back-distillation model on the image to be restored to obtain the enhanced image.

[0039] In practical applications, the student model is also the attack model. It takes the image to be restored as input and the teacher model restores it to obtain the final enhanced image. Due to the inclusion of an adaptive singular value perturbation module, the student model can only access perturbation feature maps during distillation learning. Due to the perturbation feature map The energy in the primary direction is significantly amplified, and its gradient exhibits instability, making it difficult for the student model to align with the teacher model's features, thus hindering its effective distillation learning process. Therefore, the student model is interfered with by features with added structured perturbations, failing to obtain normal representations and resulting in image recovery failure. Simultaneously, the teacher model continues inference using the unperturbed original features, maintaining high-quality output, demonstrating the ASVP module's strong protection of the teacher model. This teacher-student model architecture is a dedistillation architecture designed to prevent the student model from learning high-quality representations from the teacher model, thereby preventing the theft of teacher model knowledge.

[0040] This embodiment's solution selectively enhances the singular value range of intermediate features in the teacher model and relies on a dual-path separation mechanism to ensure that perturbations only affect the distillation channels visible to the student model. This effectively overcomes the limitations of traditional methods such as random noise and feature discarding in image restoration tasks. It not only significantly reduces the learning ability of the student model, preventing it from copying key knowledge from the teacher model, but also ensures that the structural consistency of the teacher model and the visual quality of the image are not affected during the inference phase. This fundamentally solves the technical challenge of simultaneously achieving strong perturbation strength and maintaining teacher performance in existing technologies. Through a dual-path mechanism and singular value structured perturbations, it simultaneously satisfies the technical requirements of strong distillation suppression capability and zero loss of teacher performance.

[0041] This embodiment performs system validation on several typical image restoration tasks (including super-resolution reconstruction, low-light enhancement, dehazing, deraining, and underwater image enhancement). Experimental results show that the backdistillation image restoration method based on adaptive singular value perturbation proposed in this embodiment can reduce the distillation learning ability of the student model without affecting the restoration quality of the teacher model, achieving overall performance superior to existing technologies.

[0042] To comprehensively evaluate the effectiveness of the ASVP backdistillation method in image restoration tasks, particularly its ability to inhibit student model learning while maintaining the quality of the teacher model's output, this embodiment employs several mainstream image quality evaluation metrics, including PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity Index Measure), and Learned Perceptual ImagePatch Similarity. PSNR and SSIM are used to evaluate whether the reconstruction quality of the teacher model remains stable after incorporating the ASVP module, while LPIPS measures the subjective quality consistency of the restored image based on depth-sensing features. Through comprehensive analysis of these multi-faceted metrics, the technical advantages of this embodiment's method in maintaining teacher model performance and suppressing student distillation learning ability can be fully reflected.

[0043] PSNR is a widely used full-reference quality metric for image reconstruction tasks, used to measure the pixel-level difference between a reconstructed image and a reference image. This metric is based on the assumption that the error in the pixel domain of a distorted image can be considered additive white noise, and that an image with less distortion should have a higher signal-to-noise ratio. PSNR quantifies reconstruction quality by calculating the mean-square error (MSE) of the two images and comparing it to the ratio of the maximum pixel value of the images. A higher value indicates higher image fidelity and less distortion. The calculation formula is as follows: ; in: , To reconstruct image pixel values, For reference image pixel values; The maximum possible value for a pixel (e.g., 255). A higher PSNR indicates a more consistent reconstructed image with the reference image.

[0044] SSIM is a full-reference metric that comprehensively measures image quality from three aspects: brightness, contrast, and structure. Based on the assumption that structural information in natural images is more important to the visual system, this metric considers image distortion to be primarily evaluated from the perspective of structural damage. By comparing the mean, variance, and covariance of two images within a local window, SSIM can quantify differences in brightness consistency, contrast consistency, and structural fidelity. The numerical range is [0,1], with values ​​closer to 1 indicating a sharper image and more natural structure. The calculation formula is as follows: ; in, This represents the local mean of two images; Indicates local variance; Covariance To avoid a stability constant with a denominator of zero, a higher SSIM indicates less image structural distortion and better visual consistency.

[0045] LPIPS is a perceptual similarity metric based on deep neural network features, widely used to measure the naturalness and content consistency of images at the visual perception level. Its theoretical basis is that the human visual system focuses more on the high-level semantic features of an image than merely pixel differences. LPIPS uses a deep convolutional neural network pre-trained on a large-scale image dataset to extract multi-layer features and calculates the difference distance between the reference image and the test image in the feature space of each layer to evaluate quality. The smaller the LPIPS value, the closer the image is to the reference image visually. The calculation formula is: ; in, In the first Feature maps extracted and normalized by a deep network. The layer weights, obtained through training, reflect the contribution of different layers to perceptual similarity. Because they are based on deep semantic features rather than simple pixel distance, LPIPS is more consistent with the laws of human visual perception.

[0046] Quantitative comparative analysis.

[0047] In terms of quantitative experiments, to comprehensively verify the robustness and effectiveness of the method in this embodiment under different teacher-student model structures, a knowledge distillation comparison experiment was conducted on the DIV2K dataset. The experimental setup covered various teacher networks of different sizes, including the lightweight ResNet18 (a deep residual network model with 18 weight layers), the medium-sized SwinIR (an open-source image restoration model), and the large-scale X-Restormer (an image restoration model with channel attention improvements based on Restormer (an efficient Transformer model for high-resolution image restoration)) and HAIR (a computational model commonly used for 3D hair modeling and rendering). The baselines compared with ASVP included various common distillation adversarial strategies, including noise perturbation (noise-L, noise-H), channel drop-out (dropC-L, dropC-H), and adversarial perturbation (adv-L, adv-H).

[0048] The student models included ResNet18, SwinIR, ResNet9 (a deep residual network model with 9 weight layers), and AGDN (an attention-guided distillation network for image super-resolution). The student model baseline performance, teacher model performance, and student model performance after knowledge distillation were compared and evaluated.

[0049] All methods were evaluated under the same distillation configuration and training strategy to ensure the comparability and objectivity of the results. The quantitative comparison results are shown in Table 1.

[0050] Table 1. Quantitative comparison results on the DIV2K dataset; .

[0051] As shown in Table 1, the ASVP module proposed in this embodiment exhibits significant distillation suppression effects in all teacher-student model combinations, while maintaining the original performance of the teacher model without degradation. Regarding teacher model performance, the ASVP module does not affect the reconstruction quality of the teacher model; its PSNR, SSIM, and LPIPS scores are completely consistent with the unprotected normal model, proving that this method achieves the design goal of zero damage to the teacher model.

[0052] Regarding student model performance, ASVP significantly reduces the performance of the distilled student model, with suppression effects far exceeding existing strategies such as noise perturbation, channel dropping, and adversarial perturbation. For example, when ResNet18 is used as the teacher model, the unprotected normal distillation student achieves a PSNR of 31.94 dB, while after adding ASVP, the PSNR drops significantly to 28.04 dB, and LPIPS increases from 2.098 to 5.312, indicating a significant deterioration in the perceptual quality of the student model, thus failing to effectively replicate the teacher's capabilities. Similarly, under the SwinIR teacher, the student model's LPIPS rises sharply from 1.858 to 3.858, and SSIM decreases from 0.861 to 0.612, further demonstrating that ASVP can disrupt the feature alignment process of the student network, making it difficult for the distillation objective to converge effectively.

[0053] More representative are the experimental results for the large-scale teacher models X-Restormer and HAIR. Although these high-performance models are generally easier for student networks to mimic through distillation, ASVP still maintains strong suppression capabilities, reducing student PSNR by approximately 2–3 dB and increasing LPIPS by more than 2–3 times, far exceeding other protection strategies. Among multiple teacher models, the ASVP method marked in blue achieves the strongest distillation suppression effect and reaches state-of-the-art protection performance in all five teacher-student combinations.

[0054] Overall, the quantitative results in Table 1 clearly show that: 1. The PSNR / SSIM / LPIPS of the teacher model remain completely consistent after the addition of ASVP; 2. Under different teacher structures, PSNR typically decreases by 2–4 dB, while LPIPS increases by 2–3 times, which is far superior to all noise and channel perturbation schemes; 3. Regardless of whether the teacher uses ResNet18, SwinIR, X-Restormer, or HAIR, ASVP exhibits a consistent and significant distillation suppression advantage.

[0055] Model efficiency and deployability analysis.

[0056] To further verify the computational overhead and resource consumption of the ASVP backdistillation module in a real-world deployment scenario, this paper compares several common defense methods, including none (clean), noise injection, channel dropout, and PGD-3 (anti-perturbation). The experiments measured runtime (ms / image) and peak memory (MB) at 256×256 and 4K (2160p) resolution inputs, respectively. The comparison results are shown in Table 2.

[0057] Table 2 Comparison of computational overhead and resource consumption under different strategies; .

[0058] Overall, the ASVP module in this embodiment maintains high operational efficiency and resource utilization while ensuring strong distillation protection capabilities. Specifically, the version employing full singular value perturbation (ASVP full SVD) has a runtime of 17.8 ms on a 256×256 image and 28.7 ms on a 4K image, only adding approximately 7–10 ms of overhead compared to the unprotected model, indicating that the overall computational load of this method is relatively controllable. Regarding memory usage, the peak memory usage of ASVP full SVD is 978 MB (256×256) and 2185 MB (4K), only slightly higher than the unprotected model, and still within an acceptable range.

[0059] The lightweight version of ASVP (top-40% only) proposed in this embodiment offers even greater deployment advantages. This version only amplifies the top 40% of singular values, thus significantly reducing computational load. With 256×256 and 4K input, its runtime is 12.3 ms and 19.2 ms respectively, almost identical to the unprotected model (None (clean)). Meanwhile, memory usage remains at 959 MB and 2106 MB, far lower than the 951 MB / 2138 MB and significantly higher latency (24.6 ms / 51.6 ms) of the Adversarial Perturbation-3 (PGD-3) method. This demonstrates that the lightweight version of ASVP increases computational and storage overhead almost without increasing computational and storage costs while ensuring sufficient protection capabilities, offering very high deployment flexibility.

[0060] Compared to traditional defense methods such as PGD-3, the computational overhead of ASVP in this embodiment increases by up to about 40%, but its runtime is far less than 2–3 times the latency of anti-perturbation methods, while the risk of memory overflow is significantly lower. Compared to noise injection and channel drop methods, ASVP maintains a similar level of operational efficiency, but significantly outperforms both in distillation suppression capability, demonstrating extremely high cost-effectiveness and engineering value.

[0061] In summary, this embodiment of ASVP combines high computing performance with extremely low deployment costs. In particular, the top-40% configuration can be seamlessly deployed in various real-world application scenarios without significantly impacting inference speed or GPU memory resources. Therefore, this embodiment demonstrates high practicality and scalability in cloud-based inference systems, mobile terminals, and real-time vision applications.

[0062] Example 2 This embodiment provides a back-distillation image restoration system based on adaptive singular value perturbation, including: The training module is configured to train the backdistillation model, which consists of a teacher model and a student model, using training images. The training process includes: extracting multi-scale features from the training images in parallel using both the teacher model and the student model; performing singular value decomposition on the intermediate layer features generated by the teacher model; and further processing the preceding features. The principal singular values ​​are amplified while the other singular values ​​remain unchanged, and the perturbation feature map is reconstructed in this way. The perturbation feature map is then input into the corresponding layer of the student model for feature distillation. The recovery module is configured to use a trained back-distillation model to obtain an enhanced image from the image to be recovered.

[0063] It should be noted that the above modules correspond to the steps described in Embodiment 1, and the examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should also be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0064] In further embodiments, the following is also provided: An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0065] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0066] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0067] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0068] The method in Example 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0069] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0070] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0071] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0072] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0073] Those skilled in the art will recognize that the units and algorithm steps described in connection with the various examples of this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention.

[0074] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A back-distillation image restoration method based on adaptive singular value perturbation, characterized in that, include: Training images were used to train the back distillation model, which consisted of a teacher model and a student model. The training process includes: extracting multi-scale features from the training images in parallel using both the teacher model and the student model; performing singular value decomposition on the intermediate layer features generated by the teacher model; and further processing the preceding features. The principal singular values ​​are amplified while the other singular values ​​remain unchanged, and the perturbation feature map is reconstructed in this way. The perturbation feature map is then input into the corresponding layer of the student model for feature distillation. For the image to be restored, the trained back-distillation model is used to obtain the enhanced image.

2. The back-distillation image restoration method based on adaptive singular value perturbation as described in claim 1, characterized in that, The first part of the singular value diagonal matrix obtained by singular value decomposition Each principal singular value is amplified in a controlled manner, replacing the original singular values ​​with: , This is the magnification factor. For the first One original singular value, For the first An amplified singular value.

3. The back-distillation image restoration method based on adaptive singular value perturbation as described in claim 2, characterized in that, The amplification factor is adaptively determined based on the energy distribution of the characteristic spectrum. Specifically: before calculation The cumulative energy percentage of each principal singularity Calculate the amplification factor based on the cumulative energy percentage. in, This is the adjustment coefficient.

4. The back-distillation image restoration method based on adaptive singular value perturbation as described in claim 3, characterized in that, The cumulative energy percentage is: ;in, Let t be the t-th singular value, and r be the total number of singular values.

5. The back-distillation image restoration method based on adaptive singular value perturbation as described in claim 1, characterized in that, After performing singular value amplification, use the resulting singular value amplification matrix. and the original direction matrix and Reconstructing the perturbation feature map .

6. The back-distillation image restoration method based on adaptive singular value perturbation as described in claim 1, characterized in that, The unperturbed, raw intermediate layer features are directly passed to the next network layer of the teacher model.

7. A back-distillation image restoration system based on adaptive singular value perturbation, characterized in that, include: The training module is configured to train the backdistillation model, which consists of a teacher model and a student model, using training images. The training process includes: extracting multi-scale features from the training images in parallel using both the teacher model and the student model; performing singular value decomposition on the intermediate layer features generated by the teacher model; and further processing the preceding features. The principal singular values ​​are amplified while the other singular values ​​remain unchanged, and the perturbation feature map is reconstructed in this way. The perturbation feature map is then input into the corresponding layer of the student model for feature distillation. The recovery module is configured to use a trained back-distillation model to obtain an enhanced image from the image to be recovered.

8. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the method described in any one of claims 1-6.

Images