A method and system for constructing an image restoration model based on degradation perception and feature enhancement

Abstract

This invention belongs to the field of image processing technology in computer vision and deep learning, and specifically relates to a method for constructing an image restoration model based on degradation awareness and feature enhancement. This method addresses the instability of blind restoration and the difficulty of cross-model feature alignment caused by mixed degradation such as rain, blur, and low resolution in real-world scenes. It constructs a multi-teacher-single-student knowledge transfer framework. Compared with existing purely data-driven or direct alignment restoration training methods, this invention can improve feature alignment stability and restoration quality in mixed degradation blind scenes, reduce artifacts and boundary detail loss, and reduce redundant computation through lightweight degradation representation and on-demand alignment mechanisms. It is more suitable for stable deployment and inference in edge-side resource-constrained environments.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of image processing technology in computer vision and deep learning, and particularly relates to a method and system for constructing an image restoration model based on degradation perception and feature enhancement. Background Technology

[0002] Image restoration aims to reconstruct high-quality results from degraded observation images. Typical tasks include denoising, deblurring, rain removal, and super-resolution. With the development of deep learning methods, existing techniques typically employ structures such as convolutional neural networks or Transformers to map degraded images to clear images through end-to-end supervised learning. Furthermore, to improve applicability in complex environments, some methods introduce multi-task learning frameworks, attempting to handle multiple types of degradation simultaneously using a shared backbone network. However, degradation in real-world scenarios is often not from a single source but rather a mixture of factors such as imaging links, motion, weather, and compression. Its type, intensity, and spatial distribution exhibit uncertainty and non-uniformity, making restoration models relying solely on a single degradation hypothesis or single-task data distribution prone to performance fluctuations when deployed across different scenarios.

[0003] To address the issues of hybrid degradation and cross-domain generalization, existing technologies employ approaches such as "blind recovery / degradation estimation + conditional recovery," which involves first estimating degradation information and then guiding the recovery network conditionally; or introducing knowledge distillation to transfer knowledge from multiple high-performance teacher models to a lighter student model, thus balancing recovery quality and deployment efficiency. However, these approaches generally suffer from the following shortcomings: First, the representation of degradation information is often provided in the form of global scalars or low-dimensional vectors, making it difficult to characterize the non-uniform spatial distribution of hybrid degradation, resulting in insufficient conditional guidance capabilities; second, in multi-teacher distillation scenarios, the structures and feature spaces of different teacher networks may exhibit significant heterogeneity, and scale inconsistencies, spatial misalignments, and semantic shifts can easily occur between teacher and student features, with direct alignment or simple weighted fusion potentially introducing noise and weakening effective knowledge transfer; third, some high-precision recovery models have high computational and storage overhead, making it difficult to achieve stable, real-time, or low-power deployment on resource-constrained edge devices.

[0004] In summary, the problems with existing technologies are:

[0005] (1) How to obtain a degradation characterization that reflects the mixed degradation type and spatial distribution without relying on explicit degradation labels, and use it as an effective conditional signal to stably guide the recovery process;

[0006] (2) Under the condition of multi-teacher knowledge transfer, how to achieve stable alignment and effective injection of heterogeneous features across teachers into the student feature space, and reduce the risk of misalignment and noise transmission;

[0007] (3) While ensuring the recovery effect, how to achieve robust generalization of unseen degenerate combinations through structural design and training strategies, and meet the requirements of edge deployment for computation, storage and inference stability. Summary of the Invention

[0008] The purpose of this invention is to provide an image restoration model construction method based on degradation perception and feature enhancement. This method addresses the problem of unstable blind restoration caused by uncertain intensity and uneven spatial distribution of mixed degradation types such as rain, blur, and low resolution in real-world scenes. By constructing a degradation perception module to extract degradation representation tensors that preserve spatial structure, and using these representations as conditional signals to guide feature selection and fusion in the restoration network, the method improves the restoration stability and detail fidelity under mixed degradation conditions.

[0009] Another objective is to provide a degradation condition feature enhancement projection alignment mechanism for multi-teacher knowledge transfer. This mechanism addresses the knowledge conflict and negative transfer issues caused by heterogeneous feature spaces, scale inconsistencies, and spatial misalignments in multi-teacher networks. It utilizes degradation-perceptual representations to activate and spatially enhance teacher features on demand, and achieves consistent projection alignment with the student feature space. This enables stable aggregation and effective injection of multi-source knowledge, thereby improving the convergence stability of distillation training and enhancing the generalization robustness to unseen degradation combinations.

[0010] This invention is implemented as follows: a method for constructing an image restoration model based on degradation perception and feature enhancement, the method comprising:

[0011] 1) Obtain training samples, which include an input degraded image and supervision information corresponding to the input degraded image; construct a teacher network group and a student network, wherein the teacher network group contains at least two teacher networks;

[0012] 2) Input the degraded image into the Degradation Perception Module (DPM) and extract the degradation representation tensor through the degradation encoder. ,in H and 𝑊 are the spatial dimensions of the input degraded image.

[0013] 𝑠 is the downsampling factor, and 𝐶 is the number of channels;

[0014] 3) Obtain at least two teacher intermediate features from the teacher network group and student intermediate features from the student network. Combine the teacher intermediate features, the student intermediate features, and the degenerate representation tensor. Input degradation-aware feature enhancement projector DEFEP, to in Under the conditional constraints, the intermediate features of the teacher are projected and aligned and feature enhancement is performed to obtain enhanced aligned features. The enhanced aligned features are then injected into the student network to obtain the student output.

[0015] 4) A two-stage training strategy is adopted to train the student network: in the knowledge aggregation stage, the teacher network group is used to perform knowledge transfer and constraint optimization on the student network; in the knowledge verification stage, the feature guidance of the teacher network group on the student network is cancelled, and the generalization enhancement training of the student network is carried out using hybrid degradation robustness constraints.

[0016] 5) Output the trained student network as an image restoration model.

[0017] Furthermore, the parameters of each teacher network in the teacher network group are frozen during the training of the student network, and the intermediate features of the teachers come from a preset level of feature extraction layer or feature pyramid layer.

[0018] Furthermore, the degradation encoder of the degradation perception module DPM is a lightweight convolutional residual structure, and it retains spatial location information by removing global average pooling, thereby outputting a degradation representation tensor with spatial structure. .

[0019] Furthermore, the degradation perception module DPM further includes a projection head for training, the projection head including a global pooling layer and a fully connected layer; during the training phase, a contrastive learning loss is calculated based on the projection head to learn degradation representations, and the projection head is not enabled during the inference phase.

[0020] Furthermore, the degradation-aware feature enhancement projector (DEFEP) includes:

[0021] a) Projection alignment unit, used to map the intermediate features of each teacher to a feature dimension space consistent with the intermediate features of students;

[0022] b) Degradation condition gating unit, used to determine the degradation condition based on... Generate channel selection weights and dynamically activate channels for teacher features after projection alignment;

[0023] c) Spatial enhancement unit, used to perform spatial attention enhancement on the features after channel activation, and output the enhanced aligned features.

[0024] Furthermore, the projection alignment unit includes a 1×1 convolutional mapping to match the channel dimension and performs normalization on the aligned features to improve consistency across the teacher feature space.

[0025] Furthermore, the injection method of the enhanced alignment feature is as follows: the enhanced alignment feature is fused with the student intermediate feature to obtain a fused feature, and the fused feature is used as the input of the subsequent layers of the student network, wherein the fusion includes at least one of element-wise addition and channel-wise convolutional fusion.

[0026] Furthermore, the knowledge aggregation stage optimizes at least the following loss terms:

[0027] a) Reconstruction loss, used to constrain the difference between the student output and the supervision information;

[0028] b) Structural Consistency Constraint Loss (SCR) is used to constrain the differences between student intermediate features or projected teacher features and corresponding teacher intermediate features, thereby achieving stable knowledge transfer.

[0029] Furthermore, the knowledge verification stage at least optimizes the Hybrid Degradation Robustness Constraint Loss (HCR), which is used to improve the generalization ability to unseen degradation combinations by applying hybrid degradation perturbations to the input degradation image and constraining the consistency or stability of the student network output under the condition of no teacher feature guidance.

[0030] Furthermore, the image restoration model is a multi-task restoration model, and the tasks include at least two of rain removal, deblurring, super-resolution, and denoising; the student network contains a shared feature extraction backbone and task-related output heads, and the degradation representation tensor... It serves as a conditional signal for feature fusion or feature selection in at least one task.

[0031] Another objective of this invention is to provide an image restoration model construction apparatus, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the method described herein.

[0032] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0033] Solving the technical problems of existing technologies is quite challenging: First, hybrid degradation exhibits combinatorial explosion and distribution drift characteristics, with coupling relationships between degradation factors, making it difficult to cover all degradation combinations with limited supervised samples; Second, feature heterogeneity in multi-teacher distillation stems not only from differences in network structure but also from different task preferences and representation scales, and without a reasonable alignment and selection mechanism, "knowledge conflicts" and training instability can easily occur; Third, under edge constraints, a balance needs to be struck between model capacity, inference latency, and recovery quality, avoiding both over-complexity leading to deployment difficulties and over-simplification leading to a decline in detail and boundary recovery capabilities.

[0034] Solving the above problems is of great significance: on the one hand, it can significantly improve the applicability and stability of image restoration models in real and complex environments, and reduce performance degradation caused by changes in degradation type; on the other hand, it can help realize the unified construction and efficient deployment of multi-task restoration models, providing more reliable perception and front-end enhancement capabilities for scenarios such as mobile terminals, vehicles, security and industrial vision, thereby improving the overall effect and system robustness of subsequent visual tasks such as recognition, detection and understanding.

[0035] This invention breaks through the bottleneck of hybrid degradation blind perception, achieving precise alignment and zero-conflict fusion of heterogeneous features, and significantly improving the accuracy and fidelity of image restoration in complex scenes.

[0036] (1) Technical problems and difficulties in solving them: Existing technologies face the technical challenge of unknown degradation types and extremely uneven spatial distribution when dealing with mixed degradation caused by multiple couplings such as rain, motion blur, and low resolution in real-world scenes. When using multi-teacher networks for knowledge transfer, the features extracted by different teachers exhibit severe heterogeneity in spatial distribution and semantic preferences. If only traditional linear projection or simple weighting is relied upon, it will not only fail to capture complex spatial degradation differences, but also lead to knowledge conflicts and negative transfer phenomena between different tasks, resulting in edge smoothing, structural distortion, and artifact residue in the restored image.

[0037] (2) Solution approach of the present invention: The present invention innovatively proposes a multi-task image restoration architecture based on degradation perception and feature enhancement. First, a degradation perception module (DPM) is constructed, abandoning the traditional global scalar representation, and extracting a degradation representation tensor that preserves spatial location information under an unsupervised contrastive learning mechanism. This accurately characterizes the spatial non-uniformity of hybrid degradation. Secondly, a Degradation-Aware Feature Enhancement Projector (DEFEP) is designed to... As a conditional prior, dynamic channel modulation is applied to heterogeneous features from multiple teachers, combined with an efficient spatial attention mechanism. Finally, with the support of a two-stage training strategy of knowledge aggregation and knowledge verification, the lightweight student network is guided from weighted imitation of teacher consensus to independently coping with mixed degradation perturbations, thus completely opening up the path for consistent projection alignment of cross-domain heterogeneous features to the student feature space.

[0038] (3) Creative technical effects and data support: Experimental data during the R&D process profoundly confirms the breakthrough effect of this solution: In the most challenging scenario of rain occlusion + motion blur + low resolution high-order hybrid degradation, the baseline model without the core mechanism of this invention suffers from severe feature interference, with a peak signal-to-noise ratio (PSNR) of only 26.50 dB and a structural similarity (SSIM) of only 0.810. However, after fully applying the DPM and DEFEP mechanisms of this invention, the model successfully suppressed the negative transfer between tasks, achieving an absolute PSNR gain of up to 3.65 dB (reaching 30.15 dB), and the SSIM jumped to 0.895, achieving a qualitative leap in the fidelity of high-frequency textures and underlying structures. In addition, the visualization results of latent spatial feature distribution (t-SNE) clearly show that, compared with the severe feature aliasing present in existing general restoration frameworks, this invention successfully decouples highly entangled hybrid degradation features into independent clusters with clear boundaries and tight distribution. This fully demonstrates that the proposed scheme endows the model with excellent adaptive capability to unknown degradation and robustness against interference without significantly increasing the computational overhead on the edge side.

[0039] As further supporting evidence of the inventiveness of this invention, the following important aspects are also reflected:

[0040] (1) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0041] Currently, multi-task image restoration techniques, both domestically and internationally, generally rely on non-blind assumptions about known degradation types when dealing with mixed degradation, or can only extract degradation features in the form of global scalars / low-dimensional vectors. This results in an inability to characterize the non-uniformity of degradation spatial distribution in real-world scenes. Furthermore, in the field of knowledge distillation from multi-expert models to a single lightweight model, there is a lack of a mechanism capable of dynamically aligning features based on their degradation state at the feature space level.

[0042] The technical solution of this invention fills this technological gap. This invention is the first to propose extracting a high-dimensional degenerate representation tensor that preserves spatial location information within an unsupervised contrastive learning framework. Using this as a conditional signal, pixel-level semantic navigation was achieved during multi-teacher knowledge transfer. This scheme not only fills the technical gap in accurately decoupling spatial non-uniform mixed degradation in blind restoration scenarios, but also establishes a dynamic enhancement and alignment mechanism from heterogeneous multi-source features to a unified student feature space, providing a new theoretical paradigm and engineering implementation path for all-round image restoration in complex and unknown environments.

[0043] (2) The technical solution of the present invention overcomes technical bias:

[0044] In the field of multi-teacher knowledge distillation and multi-task learning, there has long been a technical bias among those skilled in the art: that as long as the teacher model is powerful enough, the student network can naturally absorb the knowledge of all experts and achieve multi-task processing by simply projecting linear features (1x1 convolution) or by simply weighting the features of multiple teachers.

[0045] The technical solution and R&D data of this invention successfully overcome this technical bias. Extensive experiments during the R&D process confirmed that, when facing highly coupled hybrid degradation, directly sharing or averaging heterogeneous features from different teachers not only fails to improve performance but also triggers severe knowledge conflicts and negative transfer phenomena. This invention breaks the bias that blind imitation or linear averaging can achieve distillation and innovatively proposes a feature enhancement projector based on degradation-aware gating. Through nonlinear channel filtering and spatial attention enhancement using degradation priors, it proves that only by dynamically selecting and precisely aligning heterogeneous knowledge based on degradation conditions can true zero-conflict fusion of multi-task models be achieved. This breakthrough corrects a long-standing cognitive misconception in the design of multi-task networks. Attached Figure Description

[0046] Figure 1 This is an overall framework diagram provided in the embodiments of the present invention;

[0047] Figure 2 This is a structural diagram of the DEFEP module provided in an embodiment of the present invention;

[0048] Figure 3 This is a diagram of the contrastive learning mechanism provided in an embodiment of the present invention;

[0049] Figure 4 This is a diagram illustrating the feature projection process provided in an embodiment of the present invention;

[0050] Figure 5 This is a structural diagram of a degraded encoder provided in an embodiment of the present invention.

[0051] Figure 6 This is a schematic diagram of regularization comparison provided in an embodiment of the present invention.

[0052] Figure 7 This is a quantitative comparison-PSNR plot provided in an embodiment of the present invention.

[0053] Figure 8 This is a quantitative comparison-SSIM plot provided in an embodiment of the present invention.

[0054] Figure 9 This is a feature distribution visualization diagram provided in an embodiment of the present invention.

[0055] Figure 10 This is a subjective comparison chart of the GoPro dataset provided in this embodiment of the invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0057] like Figure 1 As shown, this embodiment of the invention provides a method for constructing an image restoration model based on degradation perception and feature enhancement, the method comprising:

[0058] 1) Obtain training samples, which include an input degraded image and supervision information corresponding to the input degraded image; construct a teacher network group and a student network, wherein the teacher network group contains at least two teacher networks;

[0059] 2) Input the degraded image into the Degradation Perception Module (DPM) and extract the degradation representation tensor through the degradation encoder. ,in H and 𝑊 are the spatial dimensions of the input degraded image.

[0060] 𝑠 is the downsampling factor, and 𝐶 is the number of channels;

[0061] 3) Obtain at least two teacher intermediate features from the teacher network group and student intermediate features from the student network. Combine the teacher intermediate features, the student intermediate features, and the degenerate representation tensor. Input degradation-aware feature enhancement projector DEFEP, to in Under the conditional constraints, the intermediate features of the teacher are projected and aligned and feature enhancement is performed to obtain enhanced aligned features. The enhanced aligned features are then injected into the student network to obtain the student output.

[0062] 4) A two-stage training strategy is adopted to train the student network: in the knowledge aggregation stage, the teacher network group is used to perform knowledge transfer and constraint optimization on the student network; in the knowledge verification stage, the feature guidance of the teacher network group on the student network is cancelled, and the generalization enhancement training of the student network is carried out using hybrid degradation robustness constraints.

[0063] 5) Output the trained student network as an image restoration model.

[0064] Furthermore, the parameters of each teacher network in the teacher network group are frozen during the training of the student network, and the intermediate features of the teachers come from a preset level of feature extraction layer or feature pyramid layer.

[0065] Furthermore, the degradation encoder of the degradation perception module DPM is a lightweight convolutional residual structure, and it retains spatial location information by removing global average pooling, thereby outputting a degradation representation tensor with spatial structure. .

[0066] Furthermore, the degradation perception module DPM further includes a projection head for training, the projection head including a global pooling layer and a fully connected layer; during the training phase, a contrastive learning loss is calculated based on the projection head to learn degradation representations, and the projection head is not enabled during the inference phase.

[0067] Furthermore, the degradation-aware feature enhancement projector (DEFEP) includes:

[0068] a) Projection alignment unit, used to map the intermediate features of each teacher to a feature dimension space consistent with the intermediate features of students;

[0069] b) Degradation condition gating unit, used to determine the degradation condition based on... Generate channel selection weights and dynamically activate channels for teacher features after projection alignment;

[0070] c) Spatial enhancement unit, used to perform spatial attention enhancement on the features after channel activation, and output the enhanced aligned features.

[0071] Furthermore, the projection alignment unit includes a 1×1 convolutional mapping to match the channel dimension and performs normalization on the aligned features to improve consistency across the teacher feature space.

[0072] Furthermore, the injection method of the enhanced alignment feature is as follows: the enhanced alignment feature is fused with the student intermediate feature to obtain a fused feature, and the fused feature is used as the input of the subsequent layers of the student network, wherein the fusion includes at least one of element-wise addition and channel-wise convolutional fusion.

[0073] Furthermore, the knowledge aggregation stage optimizes at least the following loss terms:

[0074] a) Reconstruction loss, used to constrain the difference between the student output and the supervision information;

[0075] b) Structural consistency constraint loss simplified regularization is used to constrain the difference between student intermediate features or projected teacher features and corresponding teacher intermediate features, thereby achieving stable knowledge transfer.

[0076] Furthermore, the knowledge verification stage at least optimizes the hybrid degradation robustness constraint loss hard regularization, which is used to improve the generalization ability to unseen degradation combinations by applying hybrid degradation perturbations to the input degradation image and constraining the consistency or stability of the student network output under the condition of no teacher feature guidance.

[0077] Furthermore, the image restoration model is a multi-task restoration model, and the tasks include at least two of rain removal, deblurring, super-resolution, and denoising; the student network contains a shared feature extraction backbone and task-related output heads, and the degradation representation tensor... It serves as a conditional signal for feature fusion or feature selection in at least one task.

[0078] Another objective of this invention is to provide an image restoration model construction apparatus, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the method described herein.

[0079] Over-introduction section:

[0080] In this specification, the image restoration task refers to restoring degraded observation images. Mapped to the restored image The degradation may include, but is not limited to, raindrops / rain streaks, motion blur or defocus blur, noise, compression artifacts, low resolution, and combinations thereof. For ease of description, let the spatial resolution of the input image be H×W, and the number of channels be 3; the features generated by the student network in the intermediate layer are denoted as... The intermediate feature generated by the k-th teacher network in the teacher network group is denoted as . .

[0081] In one embodiment, this invention employs a "multi-teacher-single-student" model, where teacher network groups provide multi-source knowledge, and student networks integrate the knowledge and output the final recovery result. To improve the stability of hybrid degradation blind recovery, this invention introduces a degradation-aware module to extract a degradation representation tensor with spatial structure from the input degradation image. Where 𝑠 is the downsampling factor (e.g., 𝑠=4), and 𝐶 is the number of representation channels. To achieve stable aggregation of heterogeneous features across teachers, this invention introduces a degenerate perceptual feature enhancement projector. Under the constraints of the conditions, the projection alignment, on-demand activation and spatial enhancement of teacher features are completed, and the enhanced knowledge is injected into the student network.

[0082] Main plan and effect description section:

[0083] The following describes an implementable process (which can be implemented by a program executed by a processor), which includes, but is not limited to, the following steps:

[0084] Step S1: Training Data and Network Construction - Obtaining Training Sample Pairs ),in For degraded images, Provide corresponding reference images or supervisory information; construct teacher network groups The student network S(⋅) is defined, where K ≥ 2. In one embodiment, the teacher network may correspond to different recovery tasks or different structural configurations to cover multiple types of degradation feature distributions; the teacher network may be a pre-trained model and its parameters may be frozen during the distillation phase.

[0085] Step S2: Degradation representation extraction of the Degradation Perception Module (DPM) will input... Input a degenerate encoder to a DPM and output a degenerate representation tensor with spatial structure. In one embodiment, the degenerate encoder may employ a lightweight residual convolutional structure and preserve spatial location information by removing global average pooling, thereby enabling... It can express the differences in mixed degradation across different spatial regions. During the training phase, to improve the discriminativeness of the degradation representation, DPM can introduce a contrastive learning branch: for different augmented views v1(⋅) and v2(⋅) of the same image, two projection vectors q1 and q2 are obtained, and the contrastive loss is minimized (example form):

[0086]

[0087] Where sim(⋅) is the similarity function and τ is the temperature coefficient. During the inference phase, the projection head can be disabled, and only the output of the degenerate encoder can be retained. .

[0088] Step S3: Acquisition of intermediate features of teachers / students Input each teacher network Tk and student network S respectively, and extract the teacher intermediate features from a preset layer (e.g., feature pyramid layer or encoder intermediate layer). Characteristics among students .

[0089] Step S4: Alignment and injection of the Degradation-Aware Feature Enhancement Projector (DE-FEP) , and The input DEFEP, in one embodiment, may include the following sub-units:

[0090] 1) Projection Alignment Unit: Maps teacher features to a feature space consistent with student features (e.g., aligning channel dimensions using 1×1 convolution / linear mapping), denoted as...

[0091] 2) Degradation condition gating unit: composed of The generation channel selection weight g is used to... Execute on-demand activation to suppress conflict knowledge (example format): Where ⊙ represents element-wise multiplication.

[0092] 3) Spatial enhancement unit: for Execution space attention enhancement is used to strengthen structural and boundary information, resulting in enhanced features. Subsequently, the enhanced features from multiple teachers were aggregated to obtain... .

[0093] 4) Injection and Fusion: [This will...] With student characteristics Fusion yields injected features The input is then fed into subsequent layers of the student network, and the restored result I is output. The fusion method can be summation or concatenated convolutions, etc. Example format:

[0094] Step S5: Two-stage training strategy (knowledge aggregation - knowledge verification) This invention adopts a two-stage training approach to balance "sufficiency of knowledge absorption" and "robustness to unseen mixed degradation".

[0095] Phase A: Knowledge Collation. In this phase, teacher-guided knowledge transfer is performed. The total loss may include reconstruction loss and structural consistency constraints (example format):

[0096] in It can be a combination of L1L_1L1 / Charbonnier / perceptual loss. This is used to constrain student features and aligned teacher features to maintain structural consistency (e.g., feature distance or structural similarity terms). , For weights.

[0097] Phase B: Knowledge Examination. In this phase, direct instruction based on teacher characteristics is eliminated, and robust reinforcement training relies solely on student networks and degenerate representations. Losses may include:

[0098] in The robustness constraint term for hybrid degradation can be achieved by applying different hybrid degradation perturbations to the same image and constraining output consistency, thereby improving the generalization ability to unseen degradation combinations.

[0099] Step S6: After training, the model outputs the student network S(⋅) as the final image reconstruction model for inference deployment. The direct effect of the above implementation process is: through By providing spatial degradation conditions, DEFEP enables on-demand alignment and enhanced injection of heterogeneous teacher knowledge. Two-stage training further improves stability under unseen mixed degradation conditions, reduces artifacts, and improves the recovery quality of boundary and texture details.

[0100] Working principle section:

[0101] (1) The mechanism of degradation perception representation: Hybrid degradation in real-world scenarios often exhibits characteristics of "type uncertainty + spatial heterogeneity". The output of DPM... Preserving spatial location can provide differentiated conditional signals for the degradation morphology of different regions, thereby enabling the student network to achieve "adaptive processing according to degradation morphology" during feature fusion and reconstruction, avoiding oversmoothing, texture mismatch and local artifact accumulation under global or unconditional conditions.

[0102] (2) The mechanism of degradation condition gating and spatial enhancement alignment: In multi-teacher distillation, different teacher features often have differences in representation scale and semantic preference. DEFEP first projects the teacher features onto a unified space, and then uses... The generation of gating weights selects knowledge relevant to the current degradation, suppressing conflicting or irrelevant features; spatial augmentation further enhances structural boundaries and detailed textures, making the injected knowledge more aligned with the current reconstruction needs of the student network, thereby improving alignment stability and reducing the risk of negative transfer.

[0103] (3) The mechanism of the two-stage training: In the knowledge aggregation stage, students can quickly absorb effective knowledge from multiple sources and establish basic recovery ability through teacher guidance; in the knowledge verification stage, after the direct supervision of teachers is cancelled, robustness constraints force students to maintain output stability under a wider range of mixed degradation perturbations, thereby improving the generalization ability to unseen degradation combinations and intensity changes, and enhancing the reliability of actual deployment.

[0104] Demonstration section (specific examples / experiments / simulations / pharmacological analysis)

[0105] To verify the effectiveness of this invention, the following experimental / simulation scheme can be adopted.

[0106] Example 1: Comparative Experiment on Restoration Quality (Subjective + Objective)

[0107] Data: Training / test sets containing multiple types of degradation and mixed degradation samples were selected; Comparison methods: Baseline A: End-to-end recovery model without degradation awareness and distillation; Comparison B: Blind recovery model with global degradation vector conditions; Comparison C: Model with multi-teacher distillation but without degradation condition alignment enhancement; Invention: DPM + DEFEP + two-stage training. Metrics: PSNR, SSIM, LPIPS (and optional edge / structure metrics).

[0108] Example 2: No generalization experiment of mixed degradation (OOD) observed

[0109] Setup: Construct degenerate combinations or intensity ranges not covered by training as OOD tests; Metrics: PSNR / SSIM / LPIPS under OOD conditions, and performance curves as a function of degradation intensity;

[0110] Example 3: Alignment and Training Stability Analysis (Mechanistic Evidence)

[0111] Analysis items: Teacher / student feature similarity before and after alignment (cosine similarity matrix or statistics), distillation / alignment loss convergence curve and fluctuation amplitude;

[0112] Example 4: Deployment and Complexity Comparison (Engineering Proof)

[0113] Metrics: number of parameters, FLOPs, peak video memory, edge inference latency / FPS (NPU / CPU / GPU logs are preferred).

[0114] I. Specific application areas or related products of this invention.

[0115] The efficient and robust image / video restoration technology of this invention can be widely and deeply applied to the following fields that are limited by complex imaging environments and edge computing power:

[0116] Smart policing and urban mobile security: Deployed on mobile terminals such as police car patrol pan-tilt units and traffic police law enforcement recorders, it can repair blurry surveillance images caused by obstruction or shaking in real time under rain, snow, low light at night, and high-speed vehicle conditions, providing high-definition visual evidence for post-event evidence collection and real-time command.

[0117] Autonomous driving and intelligent connected vehicles: As a pre-amplification module for onboard visual perception systems (such as autonomous driving domain controllers). In the face of adverse weather (rain, fog) or high-speed movement, it performs feature-level cleaning and reconstruction on degraded images captured by onboard cameras, thereby significantly improving the recognition accuracy and robustness of backend object detection (pedestrians, vehicles) and semantic segmentation algorithms.

[0118] Low-altitude drone aerial photography and reconnaissance: To address the non-uniform motion blur caused by drones during turbulent airflow and high-speed flight, as well as the low resolution problem caused by long-distance shooting, real-time super-resolution and deblurring of aerial images are achieved on the edge computing nodes mounted on the drone.

[0119] To comprehensively evaluate the overall performance of the proposed multi-task image restoration model based on degradation-aware modulation and feature-enhanced projection, a systematic comparative analysis was conducted with two mainstream methods: single-task expert models, using the teacher model that guides the student network during training as the benchmark, namely MambaIR for rain removal, GRL for motion blur removal, and PFT for super-resolution. These models were fully trained on their respective single-task datasets. For the all-around restoration model, the most representative general restoration frameworks were selected, including AirNet and AutoDIR. All comparative experiments were conducted under the same hardware environment and test dataset to ensure the fairness of the evaluation.

[0120] Figure 7 and 8The distribution of PSNR and SSIM performance for each model across four benchmarks is presented. The figures clearly show extreme performance polarization among single-task expert models. They excel in their respective training domains, but their performance drops drastically once they cross task boundaries. In the most challenging mixed degradation task, where the input image contains multiple degradation features simultaneously, single expert models are unable to cope, with PSNR generally below 24.5 dB. This confirms that single-task models alone cannot meet the complex and varied restoration requirements of the real world. While AirNet demonstrates balanced performance across tasks, its overall performance is mediocre. This is mainly because it distinguishes degradation solely through contrastive learning, lacking explicit knowledge guidance from expert networks. AutoDIR, as a state-of-the-art method based on diffusion models, outperforms AirNet, but still lags significantly behind expert models in single-task performance. This indicates that in pursuing generality, it sacrifices accuracy for some specific tasks. The proposed method demonstrates the best overall performance across all test dimensions, achieving an effective balance between cross-scene generalization and specific task accuracy. On the mixed degradation task, the method presented in this chapter achieved a highest score of 31.5 dB, thanks to the degradation-aware modulation mechanism, which accurately identifies and decouples mixed degradation features, avoiding negative transfer between tasks. In single-task tests, the performance gap between the method presented in this chapter and the single-task expert model is very close, proving that the DEFEP module effectively distills the expert knowledge of the teacher network into the student network. Quantitative experimental results show that the proposed method not only overcomes the generalization bottleneck of the single-task model, but also has a certain competitiveness in the unified restoration model. It achieves adaptive processing of multiple degradation types while maintaining a single network architecture. Further observation of the performance distribution range in the figure shows that the variance of the results of the proposed method on different test sets is smaller, effectively reducing the performance fluctuation of the model when facing extremely complex samples. In particular, its stable improvement in the SSIM index distribution further confirms that the network still has excellent low-level structure fidelity and high-frequency texture recovery ability after filtering out heavily mixed degradation artifacts.

[0121] To further explore the task decoupling mechanism within the model, this paper compares the feature distribution in the latent space of AirNet, AutoDIR, and the method presented in this chapter, and uses t-SNE technology for visualization, such as... Figure 9As shown, AirNet exhibits significant feature aliasing. Deraining and deblurring feature clusters are severely intertwined in the central region, lacking clear decision boundaries, making them prone to interference in mixed tasks. AutoDIR, a strong competitor, has successfully achieved substantial separation of different degradation features, demonstrating good task discrimination capabilities. The feature distribution of the method presented in this chapter shows superior discriminative power. Careful comparison reveals that the spacing between different task clusters in the model presented in this chapter is clearer, and the feature point distribution within clusters is more compact and convergent. This marginal improvement in boundaries and compactness demonstrates that degradation-aware modulation has higher robustness in handling complex coupled degradation features, effectively reducing ambiguous feature samples, thus supporting the leading advantage of the model presented in quantitative metrics. The visualization results are shown below. Figure 10 As shown, the input video frames (left column) exhibit noticeable blurring, edge diffusion, and local structural blurring due to camera shake and the high relative speed of the target's motion. Particularly in areas like license plate characters, tire edges, and metallic reflective surfaces, the blurring severely erases high-frequency textures, reduces local contrast, and renders details almost indistinguishable. It's important to emphasize that this video segment is from actual filming rather than artificial convolutional synthesis; therefore, its blur kernel is unpredictable and time-varying: the blur direction, intensity, and spatial distribution may change between different frames, placing high demands on the robustness of the deblurring model. Under these challenges, the model's recovery results (right column) effectively suppress nonlinear blurring caused by complex motion and maintain good visual consistency across consecutive frames. Specifically, the character outlines in the license plate area, previously blurred gray bands, converged back to clear boundaries, with more defined character spacing and stroke structure, enhancing readability. The light and shadow transitions and highlight reflections of the spare tire cover were restored, and the surface material's light and shadow layers more closely resembled the effect of a realistic, clear frame. Simultaneously, at the edges of the vehicle body and tires, the restoration results exhibited fewer common edge ghosting effects, demonstrating a relatively restrained detail enhancement process that avoided the artificial feel of over-sharpening. Furthermore, from the perspective of consecutive frames, the model in this chapter, while improving texture detail and edge sharpness, did not introduce significant cross-frame flickering or texture drift, indicating that the model not only improved single-frame clarity but also, to a certain extent, ensured temporal stability and structural consistency. Overall, these qualitative results demonstrate that the model in this chapter possesses strong adaptability to realistic, unknown, and time-varying motion blur, effectively restoring key detail areas.

[0122] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0123] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for constructing an image restoration model based on degradation perception and feature enhancement, characterized in that, The method includes: 1) Obtain training samples, which include an input degraded image and supervision information corresponding to the input degraded image; construct a teacher network group and a student network, wherein the teacher network group contains at least two teacher networks; 2) Input the degraded image into the Degradation Perception Module (DPM) and extract the degradation representation tensor through the degradation encoder. ,in H and 𝑊 are the spatial dimensions of the input degraded image. 𝑠 is the downsampling factor, and 𝐶 is the number of channels; 3) Obtain at least two teacher intermediate features from the teacher network group and student intermediate features from the student network. Combine the teacher intermediate features, the student intermediate features, and the degenerate representation tensor. Input degradation-aware feature enhancement projector DEFEP, to in Under the conditional constraints, the intermediate features of the teacher are projected and aligned and feature enhancement is performed to obtain enhanced aligned features. The enhanced aligned features are then injected into the student network to obtain the student output. 4) A two-stage training strategy is adopted to train the student network: in the knowledge aggregation stage, the teacher network group is used to perform knowledge transfer and constraint optimization on the student network; in the knowledge verification stage, the feature guidance of the teacher network group on the student network is cancelled, and the generalization enhancement training of the student network is carried out using hybrid degradation robustness constraints. 5) Output the trained student network as an image restoration model.

2. The image restoration model construction method based on degradation perception and feature enhancement according to claim 1, characterized in that, The parameters of each teacher network in the teacher network group are frozen during the training of the student network, and the intermediate features of the teachers come from the feature extraction layer or feature pyramid layer of the preset level.

3. The image restoration model construction method based on degradation perception and feature enhancement according to claim 1, characterized in that, The degradation perception module (DPM) employs a lightweight convolutional residual structure as its degradation encoder. It removes global average pooling to preserve spatial location information, thereby outputting a degradation representation tensor with spatial structure. .

4. The method for constructing an image restoration model based on degradation perception and feature enhancement according to claim 1 or 3, characterized in that, The degradation awareness module (DPM) further includes a projection head for training, which includes a global pooling layer and a fully connected layer; during the training phase, a contrastive learning loss is calculated based on the projection head to learn degradation representations, and the projection head is not enabled during the inference phase.

5. The method for constructing an image restoration model based on degradation perception and feature enhancement according to claim 1, characterized in that, The degradation-sensory feature enhancement projector (DEFEP) includes: a) Projection alignment unit, used to map the intermediate features of each teacher to a feature dimension space consistent with the intermediate features of students; b) Degradation condition gating unit, used to determine the degradation condition based on... Generate channel selection weights and dynamically activate channels for teacher features after projection alignment; c) Spatial enhancement unit, used to perform spatial attention enhancement on the features after channel activation, and output the enhanced aligned features.

6. The method for constructing an image restoration model based on degradation perception and feature enhancement according to claim 5, characterized in that, The projection alignment unit contains a 1×1 convolutional mapping to match the channel dimension and performs normalization on the aligned features to improve consistency across the teacher feature space.

7. The method for constructing an image restoration model based on degradation perception and feature enhancement according to claim 1, characterized in that, The method of injecting the enhanced alignment feature is as follows: the enhanced alignment feature is fused with the student intermediate feature to obtain a fused feature, and the fused feature is used as the input of the subsequent layers of the student network, wherein the fusion includes at least one of element-wise addition and channel-wise convolutional fusion.

8. The method for constructing an image restoration model based on degradation perception and feature enhancement according to claim 1, characterized in that, The knowledge aggregation phase optimizes at least the following loss terms: a) Reconstruction loss, used to constrain the difference between the student output and the supervision information; b) Structural Consistency Constraint Loss (SCR) is used to constrain the differences between student intermediate features or projected teacher features and corresponding teacher intermediate features, thereby achieving stable knowledge transfer.

9. The method for constructing an image restoration model based on degradation perception and feature enhancement according to claim 1, characterized in that, The knowledge verification phase optimizes at least the Hybrid Degradation Robustness Constraint Loss (HCR), which is used to improve the generalization ability to unseen degradation combinations by applying hybrid degradation perturbations to the input degradation image and constraining the consistency or stability of the student network output under conditions without teacher feature guidance.

10. The method for constructing an image restoration model based on degradation perception and feature enhancement according to claim 1, characterized in that, The image restoration model is a multi-task model, and the tasks include at least two of rain removal, deblurring, super-resolution, and denoising; the student network contains a shared feature extraction backbone and task-related output heads, and the degradation representation tensor... It serves as a conditional signal for feature fusion or feature selection in at least one task.

Images