CN-121981892-A - Single image super-resolution oriented stable feature enhancement method, device and equipment

CN121981892ACN 121981892 ACN121981892 ACN 121981892ACN-121981892-A

Abstract

The invention relates to a method, a device and equipment for enhancing stable characteristics of super-resolution of a single image, belonging to the technical field of image processing. The method comprises the steps of mapping a low-resolution image to a high-dimensional feature space to obtain initial high-dimensional features, constructing a deep backbone network comprising a plurality of parallel mixed enhancement modules, carrying out feature enhancement processing on the initial high-dimensional features, carrying out local stable propagation processing of self-adaptive subspace division on the initial high-dimensional features in each mixed enhancement module, carrying out global priori injection processing with a learnable prototype library, carrying out residual fusion on the global enhancement features and the initial high-dimensional features, outputting current mixed enhancement features, carrying out residual fusion on the parallel mixed enhancement features, carrying out residual aggregation on the parallel mixed enhancement features and the initial high-dimensional features, and carrying out up-sampling and reconstruction processing to output a final super-resolution result. The invention gives consideration to local detail restoration and full-length Cheng Yilai modeling of the image, and improves the feature expression stability and reconstruction robustness.

Inventors

ZHU XIANGYUAN
LIU XUCHONG
SU XIN
SHEN FUHUI
HUANG WEI

Assignees

湖南警察学院

Dates

Publication Date: 20260505
Application Date: 20260402

Claims (10)

1. A method for enhancing a stable characteristic of super-resolution of a single image, the method comprising: Mapping the low-resolution image to a high-dimensional feature space to obtain initial high-dimensional features; constructing a deep backbone network, wherein the deep backbone network comprises a plurality of parallel mixed enhancement modules, and the initial high-dimensional characteristics are subjected to characteristic enhancement processing through the mixed enhancement modules to obtain a plurality of parallel mixed enhancement characteristics; in each mixed enhancement module, carrying out local stable propagation processing of self-adaptive subspace division on the initial high-dimensional features to obtain local enhancement features, carrying out global priori injection processing with a learnable prototype library on the local enhancement features to obtain global enhancement features, carrying out residual fusion on the global enhancement features and the initial high-dimensional features, and outputting current mixed enhancement features; fusing all the parallel mixed enhancement features and then carrying out residual aggregation on the fused parallel mixed enhancement features and the initial high-dimensional features to obtain high-quality reconstruction features; And mapping the high-quality reconstruction features to a target high-resolution space through upsampling and reconstruction processing, and outputting a final super-resolution result.
2. The method for enhancing the stable characteristics of the super resolution of the single image according to claim 1, wherein the performing the local stable propagation processing of the adaptive subspace division on the initial high-dimensional characteristics to obtain the local enhanced characteristics comprises: dividing the initial high-dimensional features into a plurality of subspace features which are not overlapped with each other along a channel dimension; In each subspace feature, generating an adaptive base matrix by learning an adaptive low-rank generation base, and projecting the features in the group to obtain a corresponding low-dimensional coefficient; Performing stable propagation updating on the low-dimensional coefficient in a low-dimensional coefficient domain to obtain an updated low-dimensional coefficient; The updated low-dimensional coefficients are put back to the corresponding subspace features according to the self-adaptive base matrix, and the reconstructed subspace features are obtained; and splicing all the reconstructed subspace features along the channel dimension to obtain the local enhancement features.
3. The method for enhancing the super-resolution stable characteristics of a single image according to claim 2, wherein the characteristics in the group are as follows: ; In the formula, Represent the first A subspace feature; represent the first An adaptive base matrix corresponding to each subspace feature; represent the first And the low-dimensional coefficients corresponding to the subspace features.
4. The method for enhancing the super-resolution stable characteristics of a single image according to claim 3, wherein the low-dimensional coefficients are stably propagated and updated in a low-dimensional coefficient domain, and an update formula is as follows: ; In the formula, Represent the first Updated low-dimensional coefficients corresponding to the subspace features; represent the first Propagation intensity coefficients corresponding to the subspace features; represent the first Propagation weights of the individual subspaces in the coefficient domain.
5. The method for enhancing the super-resolution stable characteristics of a single image according to any one of claims 2 to 4, further comprising, after obtaining the reconstructed subspace features, before stitching along the channel dimension: Extracting a global description vector of the reconstructed subspace feature for each subspace feature, and generating a scaling item and a biasing item according to the global description vector; carrying out statistical modulation on the corresponding reconstructed subspace characteristics according to the scaling items and the bias items to obtain modulated subspace characteristics; And splicing all the modulated subspace features along the channel dimension to obtain the local enhancement features.
6. The method for enhancing the super-resolution stable characteristics of a single image according to claim 5, wherein the corresponding reconstructed subspace characteristics are statistically modulated according to the scaling term and the bias term, and the process expression is as follows: ; In the formula, Represent the first The modulated subspace features; represent the first Reconstructing subspace features; Representing a scaling function; representing a bias function; representing a global description vector; Representing element-by-element multiplication.
7. The method for enhancing a stable feature for super resolution of a single image according to any one of claims 1 to 4, wherein performing global prior injection processing with a learnable prototype library on the local enhanced feature to obtain a global enhanced feature comprises: respectively carrying out low-dimensional embedding and normalization processing on the local enhancement features and the learnable prototype library to obtain feature embedding and prototype embedding with uniform dimensions; Constructing a route cost matrix based on the feature embedding and prototype embedding, and performing entropy regularization and route optimization of balance constraint on the route cost matrix to obtain a balance route matrix; performing spectrum smoothing processing on the learnable prototype library to obtain a smoothed structure prototype library; carrying out weighted aggregation on the smoothed structure prototype library according to the balanced routing matrix to generate a global priori feature; and carrying out self-adaptive gating fusion on the local enhancement features and the global priori features to obtain the global enhancement features.
8. The single image super-resolution oriented stable feature enhancement method according to claim 7, wherein the local enhancement features and the global prior features are adaptively gated and fused, and the expression is: ; In the formula, Representing global enhancement features; representing local enhancement features; representing global a priori features; representing a gating vector; Representing a learnable output mapping matrix; Representing element-by-element multiplication.
9. A single image super-resolution oriented stable feature enhancement device, the device comprising: the shallow feature extraction module is used for mapping the low-resolution image to a high-dimensional feature space to obtain an initial high-dimensional feature; the deep layer feature enhancement module is used for constructing a deep layer backbone network, the deep layer backbone network comprises a plurality of parallel mixed enhancement modules, the feature enhancement processing is respectively carried out on the initial high-dimensional features through the mixed enhancement modules to obtain a plurality of parallel mixed enhancement features, in each mixed enhancement module, the local stable propagation processing of self-adaptive subspace division is carried out on the initial high-dimensional features to obtain local enhancement features, the global priori injection processing with a leachable prototype library is carried out on the local enhancement features to obtain global enhancement features, and residual fusion is carried out on the global enhancement features and the initial high-dimensional features to obtain current mixed enhancement features; the feature fusion aggregation module is used for fusing all the parallel mixed enhancement features and then aggregating the fused features with the initial high-dimensional feature residual errors to obtain high-quality reconstruction features; And the up-sampling reconstruction output module is used for mapping the high-quality reconstruction features to a target high-resolution space through up-sampling and reconstruction processing and outputting a final super-resolution result.
10. A computer device comprising a memory and a processor, characterized in that the memory stores a computer program, which when executed implements the steps of the single image super-resolution oriented stable feature enhancement method according to any of claims 1 to 8.

Description

Single image super-resolution oriented stable feature enhancement method, device and equipment Technical Field The invention relates to the technical field of image processing, in particular to a method, a device and equipment for enhancing stable characteristics of super-resolution of a single image. Background The single image super Resolution (SingleImageSuper-Resolution, SISR) aims to recover high Resolution results from a single low Resolution image, which is a typical problem of the inverse of discomfort. Because the low-resolution observation loses a large amount of high-frequency texture and edge details in the downsampling process, the same low-resolution input often corresponds to a plurality of possible high-resolution solutions, so how to accurately recover structural information, texture details and visual fidelity under limited observation conditions is always a core research direction in the field of image reconstruction. Due to the development of deep learning, the super-resolution method based on the convolutional neural network has remarkable progress in reconstruction quality and is widely applied to scenes such as medical imaging, image acquisition, video monitoring, mobile vision enhancement and the like. Early convolution-based super-resolution methods, typically improved reconstruction performance by stacking deeper networks, designing residual connections, or multi-scale feature fusion strategies. The method has strong capability in the aspect of local texture recovery, is limited by a receptive field and a characteristic modeling mode of local convolution, and is difficult to fully describe the long-range dependency relationship and cross-region structural similarity commonly existing in the image. To address this problem, researchers have introduced structures such as attention mechanisms, transformers, and state space modeling into super-resolution tasks to enhance global context modeling capabilities. However, when the method performs dense interaction in the complete high-dimensional feature space, higher computational complexity is easily introduced, and meanwhile, problems of redundancy correlation accumulation, local disturbance amplification, unstable feature expression and the like can be possibly brought. Particularly in a light super-resolution scene, if explicit constraints on the characteristic interaction path and the structure propagation process are lacking, the model can easily obtain a larger modeling range and sacrifice numerical stability and reconstruction robustness. Disclosure of Invention Based on the above, it is necessary to provide a method, a device and equipment for enhancing stable characteristics of super-resolution of single image, which can realize local texture restoration and effective modeling of long-range dependence of the whole situation and simultaneously consider characteristic stability and reconstruction robustness. A method for enhancing a stable feature for super resolution of a single image, the method comprising: Mapping the low-resolution image to a high-dimensional feature space to obtain initial high-dimensional features; constructing a deep backbone network, wherein the deep backbone network comprises a plurality of parallel mixed enhancement modules, and the initial high-dimensional characteristics are subjected to characteristic enhancement processing through the mixed enhancement modules to obtain a plurality of parallel mixed enhancement characteristics; in each mixed enhancement module, carrying out local stable propagation processing of self-adaptive subspace division on the initial high-dimensional features to obtain local enhancement features, carrying out global priori injection processing with a learnable prototype library on the local enhancement features to obtain global enhancement features, carrying out residual fusion on the global enhancement features and the initial high-dimensional features, and outputting current mixed enhancement features; fusing all the parallel mixed enhancement features and then carrying out residual aggregation on the fused parallel mixed enhancement features and the initial high-dimensional features to obtain high-quality reconstruction features; And mapping the high-quality reconstruction features to a target high-resolution space through upsampling and reconstruction processing, and outputting a final super-resolution result. On the other hand, still provide a stable characteristic enhancement device towards single image super resolution, said device includes: the shallow feature extraction module is used for mapping the low-resolution image to a high-dimensional feature space to obtain an initial high-dimensional feature; The deep layer feature enhancement module is used for constructing a deep layer backbone network, the deep layer backbone network comprises a plurality of parallel mixed enhancement modules, the feature enhancement processing is respectively carried out on the initial hi