Instrument image enhancement method combining low light enhancement and deblurring

Abstract

The invention provides an instrument image enhancement method combining low light enhancement and deblurring, and relates to the technical field of image processing. The method comprises the following steps: establishing an image enhancement network for instrument image enhancement, establishing a prompt contrast learning mechanism based on CLIP semantic guidance, and training the image enhancement network; in combination with two tasks of low-light enhancement and deblurring, the brightness, structure and semantic information of the image are recovered at the same time, so that the quality and recognition performance of the image are effectively improved. According to the method, the problems of detail loss, color shift and artifacts caused by staged independent processing in a traditional method are effectively avoided; the method is obviously superior to an existing algorithm in the aspects of brightness balance, marginal definition and texture fidelity, the enhanced image does not have obvious artifacts or color cast, and the visual quality is obviously improved.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and particularly relates to an instrument image enhancement method that combines low-light enhancement and deblurring. Background Technology

[0002] In the field of image processing, numerous studies have proposed corresponding technical solutions for the enhancement and restoration of low-light and blurred images. Traditional low-light enhancement methods, such as those based on Retinex theory and histogram equalization, mainly improve brightness by separating the reflection and illumination components of the image. However, these methods often introduce artifacts or texture distortion in low-light environments, affecting the final visual effect. Regarding deblurring, existing algorithms typically rely on convolution kernel estimation or prior constraint optimization. Under low-light conditions or severe motion blur, the estimation of the blur kernel is often unstable, leading to over-sharpening or edge breakage, making it difficult for the image quality to meet practical application requirements.

[0003] With the development of deep learning technology, end-to-end image enhancement networks have been gradually applied, especially in the processing of low light and blur. However, most current deep learning models can only optimize for a single degradation problem (low light or blur), making it difficult to effectively handle multiple degradation problems within a unified framework. Therefore, how to simultaneously improve the brightness, sharpness, and structural information of an image within the same network framework remains a challenge in the field of image enhancement.

[0004] Existing low-light enhancement and deblurring methods generally suffer from the following defects: (1) Limitations of low-light enhancement methods: Traditional low-light enhancement techniques, such as Retinex theory and histogram equalization, can improve the brightness of images, but under noise interference or blurring conditions, they often lead to image artifacts, color distortion or loss of details. Especially when low light and motion blur coexist, these methods cannot effectively handle multiple degradations of the image, ultimately resulting in unsatisfactory image quality. (2) Instability of deblurring algorithms: Deblurring algorithms usually rely on the estimation of the blur kernel, but in low-light environments, the estimation of the blur kernel is often unstable, easily causing problems such as over-sharpening and edge breakage, resulting in poor image quality. In addition, most deblurring methods often only focus on the local features of the image during the optimization process, ignoring the recovery of global illumination and structural information, thus affecting the overall image quality. (3) Inability to uniformly handle the dual degradation of low light and blur: Existing methods often treat low-light enhancement and deblurring as two independent processing steps, which often leads to error accumulation during the processing, resulting in poor final image quality. Meanwhile, many deep learning models only optimize for a single degradation problem and lack the ability to handle low light and blur simultaneously. Therefore, the inability to perform joint low light and blur recovery within the same framework limits their application in complex scenarios. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a combined low-light enhancement and deblurring method for instrument images, aiming to solve the problem that low-light enhancement and deblurring processes cannot be uniformly and collaboratively optimized in existing technologies. Specifically, this invention designs an integrated network framework that combines the two tasks of low-light enhancement and deblurring, simultaneously restoring the brightness, structure, and semantic information of an image within a unified framework, thereby effectively improving image quality and recognition performance.

[0006] In a first aspect, the present invention provides a method for enhancing instrument images by combining low-light enhancement and deblurring, comprising the following steps:

[0007] An image enhancement network for instrument image enhancement is constructed to perform low-light enhancement and deblurring on low-light blurred instrument images. The network includes a decomposition module, a low-light enhancement encoder, multiple fusion layers, a deblurring decoder, a convolutional mapping layer, and an inverse HVI transform module.

[0008] A CLIP-based semantic guidance-guided prompting contrast learning mechanism is established to train the image enhancement network;

[0009] Obtain the low-light blurred instrument image to be enhanced, and perform low-light enhancement and deblurring processing on the low-light blurred image to be enhanced based on the trained image enhancement network to generate the enhanced instrument image;

[0010] The decomposition module is used to decompose the input low-light blurred image into chroma and luminance, decomposing the low-light blurred instrument image into a luminance map and a chroma map; the low-light enhancement encoder is used to jointly enhance the luminance and chroma of the luminance map and chroma map, generating a denoised chroma feature map and an enhanced luminance feature map; the fusion layer is used to fuse the denoised chroma feature map and the enhanced luminance feature map into a fused enhanced feature map; the deblurring decoder is used to perform layer-by-layer deblurring and thinning reconstruction on the fused enhanced feature map, generating a preliminary reconstructed image; the convolutional mapping layer and the inverse HVI transform module are used to perform color space conversion on the preliminary reconstructed image to obtain the enhanced instrument image.

[0011] Furthermore, the specific method for decomposing the low-light blurred instrument image into a luminance map and a chrominance map is as follows:

[0012] A luminance map is generated based on the maximum value of each pixel in the three RGB channels of the low-light blurred image.

[0013] A saturation map and a preliminary chromaticity map are generated based on the luminance map. The chromaticity component of each pixel in the preliminary chromaticity map is mapped to two polarized orthogonal components.

[0014] A first adaptive light intensity compression function is established based on the brightness map. The first adaptive light intensity compression function and the saturation map are multiplied element-wise by the two polarized orthogonal components to obtain two intermediate orthogonal components.

[0015] By splicing the two intermediate orthogonal components along the channel dimension, a chromaticity map is obtained.

[0016] Furthermore, the low-light enhancement encoder includes multiple cross-attention layers;

[0017] The specific method for joint enhancement of luminance and chrominance in low-light enhancement encoders is as follows:

[0018] The luminance and chrominance maps are input into the low-light enhancement encoder and passed through multiple cross-attention layers to generate a luminance-equalized image.

[0019] Each cross-attention layer includes a multi-head cross-attention module, a brightness enhancement module, and a chroma denoising module;

[0020] The multi-head cross-attention module includes a chroma branch and a luminance branch, which are used to process the chroma map and luminance map respectively, extract chroma features and luminance features, input the chroma features into the chroma denoising module, and input the luminance features into the luminance enhancement module.

[0021] The chroma denoising module comprises multiple convolutional sub-modules. The chroma features input to the chroma denoising module are processed through the first convolutional sub-module and then fed into the second and third convolutional sub-modules. The outputs of the second and third convolutional sub-modules are non-linearly mapped using the Tanh activation function, and then their residuals are added to the output of the first convolutional module to obtain two chroma enhancement features. An element-wise multiplication fusion operation is performed on the two chroma enhancement features, and the fused features are input to the fourth convolutional sub-module. The output of the fourth convolutional sub-module is then added to the residuals of the input chroma features to obtain a denoised chroma feature map.

[0022] The brightness enhancement module includes multiple convolutional sub-modules. The brightness features input to the brightness enhancement module are processed through the fifth convolutional sub-module and then input into the sixth and seventh convolutional sub-modules, respectively. The outputs of the sixth and seventh convolutional sub-modules are nonlinearly mapped using the Tanh activation function, and then the residuals are added to the output of the fifth convolutional module to obtain two intermediate brightness features. After performing an element-wise multiplication fusion operation on the two intermediate brightness features, the fused features are input into the eighth convolutional sub-module. The output of the eighth convolutional sub-module is added to the residuals of the brightness features input to the brightness enhancement module to obtain the enhanced brightness feature map.

[0023] Furthermore, the fusion layer includes multiple convolutional modules and a multi-stage compressed activation fusion module;

[0024] The specific method for fusing the denoised chromaticity feature map and the enhanced luminance feature map into a fused enhanced feature map is as follows:

[0025] Two convolutional modules are used to convolve the denoised chroma feature map and the enhanced luminance feature map respectively. The enhanced luminance feature after convolution is used to linearly weight the denoised chroma feature after convolution to obtain the preliminary enhanced chroma feature.

[0026] The initial enhanced chroma features are input into the multi-stage compression activation fusion module to obtain the final enhanced chroma features;

[0027] The final enhanced chroma feature and the convolutional enhanced luminance feature map are concatenated by channel and passed through a convolution module to obtain the fused enhanced feature map.

[0028] Furthermore, the deblurring decoder includes multiple fuzzy perception reconstruction layers; each fuzzy perception reconstruction layer includes a multi-scale fuzzy perception residual module and a wavelet reconstruction module, with each multi-scale fuzzy perception residual module connected to a wavelet reconstruction module; the multi-scale fuzzy perception residual module includes multiple depthwise separable convolutional sub-modules, a multi-scale feature extraction branch, and a fuzzy attention perception branch.

[0029] The specific method for layer-by-layer deblurring and thinning reconstruction of the fused and enhanced feature map is as follows:

[0030] Multiple fused and enhanced feature maps output from multiple fusion layers are input into the deblurring decoder, and after passing through multiple blur-aware reconstruction layers, a preliminary reconstructed image is generated, including:

[0031] The fused enhanced feature map and the feature map output from the previous fuzzy perception reconstruction layer are used as input, or only the feature map output from the previous fuzzy perception reconstruction layer is used as input, and then passed through the multi-scale fuzzy perception residual module and the wavelet reconstruction module in sequence.

[0032] In the multi-scale fuzzy perception residual module, a first depthwise separable convolutional submodule is used to process the feature map input to the multi-scale fuzzy perception residual module. The first depthwise separable convolutional submodule is processed, comprising a 3×3 depthwise separable convolution DSConv kernel and a LeakyReLU activation function; the feature map output by the first depthwise separable convolutional submodule is processed. Input the multi-scale feature extraction branch and the fuzzy attention perception branch respectively; input the features output by the multi-scale feature extraction branch. Fuzzy Enhancement Features of Fuzzy Attention Perception Branch Output The splicing is performed along the channel dimension to obtain the splicing features. ; splicing features The input is fed into a second-depth separable convolutional submodule, where feature integration and channel compression are performed to obtain intermediate fused features. ; integrate intermediate features Feature maps of the input multi-scale fuzzy perception residual module Establish cross-branch residual connections to obtain intermediate enhanced feature maps from the output of the multi-scale fuzzy perception residual module (MBR). .

[0033] intermediate enhanced feature map The input is fed into the wavelet reconstruction module to generate a preliminary reconstructed image. .

[0034] Furthermore, the multi-scale feature extraction branch includes multiple parallel depth-wise dilated convolution submodules, depth-separable dilated convolution submodules, and channel attention submodules;

[0035] Multiple parallel depthwise dilated convolutional submodules are used to extract feature maps from the input multi-scale feature branches. Convolution is performed, and the output features of multiple parallel depthwise dilated convolution sub-modules are concatenated along the channel dimension to obtain multi-scale fused features. ;

[0036] Multi-scale feature fusion using depthwise separable dilated convolution submodules Perform compression mapping to obtain fused features. ;

[0037] Fusion features Input is fed into the channel attention submodule, and the output is attention-enhanced features. ;

[0038] Attention-enhancing features Feature maps of the input multi-scale feature extraction branch The residual connections are formed by element-wise addition, resulting in the features output by the multi-scale feature extraction branch. .

[0039] Furthermore, the fuzzy attention perception branch includes multiple strip pooling operations of different scales, two 1×1 convolutional sub-modules, and two 3×3 convolutional sub-modules; the first 1×1 convolutional sub-module includes a 1×1 convolution and a ReLU activation function, and the second 1×1 convolutional sub-module includes a 1×1 convolution and a Sigmoid activation function; the first 3×3 convolutional sub-module includes a 3×3 convolution and a ReLU activation function, and the second 3×3 convolutional sub-module includes a 3×3 convolution and a Sigmoid activation function.

[0040] Strip pooling operations of different scales were used to refine the feature maps of the fuzzy attention perception branch along the horizontal and vertical directions, respectively. Directional fuzzy information is extracted to obtain fuzzy response tensors at different scales;

[0041] The fuzzy response tensors at all scales are concatenated and input into two 1×1 convolutional submodules for processing. The features output from the second 1×1 convolutional module are then combined with the feature maps input to the fuzzy attention perception branch. Element-level weighted fusion is performed to obtain intermediate features ;

[0042] Use two 3×3 convolutional submodules for intermediate features The fuzzy attention mask output by the second 3×3 convolutional submodule is processed. and intermediate features Element-level weighted fusion is performed to obtain fuzzy enhanced features. .

[0043] Furthermore, the wavelet reconstruction module includes a discrete wavelet transform unit and a discrete wavelet inverse transform unit;

[0044] Generate final enhanced output features The specific method is as follows:

[0045] intermediate enhanced feature map The input is fed into the wavelet reconstruction module, which uses Haar wavelets to enhance the intermediate feature maps. Decomposed into the wavelet frequency domain, sub-band feature maps of different frequencies are generated, including low-frequency sub-band feature maps, horizontal high-frequency sub-band feature maps, vertical high-frequency sub-band feature maps, and diagonal high-frequency sub-band feature maps;

[0046] Subband feature maps of different frequencies are input into the feature recovery network for independent reconstruction, resulting in multiple reconstructed subband feature maps, including reconstructed low-frequency subband feature maps, reconstructed horizontal high-frequency subband feature maps, reconstructed vertical high-frequency subband feature maps, and reconstructed diagonal high-frequency subband feature maps.

[0047] Multiple reconstructed sub-band feature maps are input into the discrete wavelet inverse transform unit to obtain a preliminary reconstructed image. .

[0048] Furthermore, the specific method for color space conversion of the initially reconstructed image is as follows:

[0049] The initially reconstructed image is sequentially input into a convolutional mapping layer and an inverse HVI transform module. The convolutional mapping layer outputs a three-channel feature representation of the initially reconstructed image, and the inverse HVI transform module decomposes the three-channel feature representation into two orthogonal chromaticity components. , And a brightness component;

[0050] A second adaptive light intensity compression function is constructed based on the luminance component, and this function is then used to compress the two orthogonal chrominance components. and Normalization and scaling are performed to obtain the corrected chromaticity components. and ;

[0051] Corrected chromaticity components , The luminance component is mapped to the HSV color space. Based on the standard HSV to RGB segmented mapping rules, the resulting HSV color space representation is converted to the sRGB color space, and the enhanced instrument image is output.

[0052] Furthermore, a CLIP-based semantic-guided cue contrast learning mechanism is established, and the specific method for training the image enhancement network is as follows:

[0053] Construct a pair of learnable text cue vectors to represent different image quality states. The learnable text cue vectors include positive cue and negative cue. The positive cue is initialized with a semantic embedding describing a clear instrument image in good light, and the negative cue is initialized with a semantic embedding describing a blurry instrument image in low light.

[0054] The CLIP model is used to iteratively optimize positive and negative cues. Low-light blurred images and reference well-lit clear images are input into the CLIP model's image encoder to extract latent semantic features. Positive and negative cues are input into the CLIP text encoder to extract semantic embedding representations. A cue contrast loss function under illumination and blur conditions is constructed. By minimizing the cue contrast loss function, gradient updates are performed on the positive and negative cues.

[0055] The iteratively optimized positive and negative cues are fixed and embedded into the training phase of the image augmentation network. A CLIP-based augmentation loss is established for training the image augmentation network. The trained image augmentation network is obtained by minimizing the CLIP-based augmentation loss.

[0056] Secondly, this application proposes an electronic device comprising: one or more processors, and a memory for storing instructions that, when executed by the one or more processors, cause the one or more processors to perform the combined low-light enhancement and deblurring instrument image enhancement method.

[0057] Thirdly, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the aforementioned instrument image enhancement method combining low-light enhancement and deblurring.

[0058] Fourthly, this application proposes a computer program product, including a computer program or instructions that, when executed by a processor, implement the aforementioned method for combined low-light enhancement and deblurring of instrument images.

[0059] The beneficial effects of adopting the above technical solution are as follows: The instrument image enhancement method provided by this invention, which combines low-light enhancement and deblurring, can significantly improve the brightness, sharpness, and structural integrity of images under complex imaging conditions such as low light and motion blur. By jointly optimizing brightness restoration and blur removal within a unified framework, this invention effectively avoids the problems of detail loss, color shift, and artifacts caused by independent processing in stages in traditional methods; it is significantly superior to existing algorithms in terms of brightness uniformity, edge sharpness, and texture fidelity, and the enhanced image has no obvious artifacts or color cast, resulting in a significant improvement in subjective visual quality. Attached Figure Description

[0060] Figure 1 A schematic diagram of the deep learning network framework structure for instrument image enhancement provided in this embodiment of the invention;

[0061] Figure 2 A schematic diagram of the low-light enhancement encoder structure provided in this embodiment of the invention;

[0062] Figure 3 A schematic diagram of the deblurring decoder structure provided in this embodiment of the invention;

[0063] Figure 4 A schematic diagram of the wavelet reconstruction module structure provided in this embodiment of the invention. Detailed Implementation

[0064] The specific implementation methods of this application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0065] Example 1:

[0066] A method for enhancing instrument images by combining low-light enhancement and deblurring includes the following steps:

[0067] Construct an image enhancement network for instrument image enhancement, such as Figure 1 As shown, it includes a decomposition module, a low-light enhancement encoder, multiple fusion layers, a deblurring decoder, a convolutional mapping layer, and an inverse HVI transform module; the image enhancement network is used to perform low-light enhancement and deblurring on low-light blurred instrument images to generate enhanced instrument images;

[0068] The decomposition module decomposes the input low-light blurred image into chroma and luminance maps, separating the low-light blurred instrument image into a luminance map and a chroma map. The low-light enhancement encoder jointly enhances the luminance and chroma of the preprocessed instrument image, generating a denoised chroma feature map and an enhanced luminance feature map. The fusion layer fuses the denoised chroma feature map and the enhanced luminance feature map into a fused enhanced feature map. The deblurring decoder performs layer-by-layer deblurring and thinning reconstruction on the fused enhanced feature map, generating a preliminary reconstructed image. The convolutional mapping layer and the inverse HVI transform module perform color space conversion on the preliminary reconstructed image to obtain the enhanced instrument image.

[0069] The decomposition module performs chromaticity and luminance decomposition processing on the input low-light blurred image, decomposing the low-light blurred image into a luminance image and a chromaticity image. The luminance image is used to enhance brightness, and the chromaticity image is used for color correction.

[0070] The specific method for decomposing a low-light blurred instrument image into a luminance map and a chrominance map is as follows:

[0071] Based on Max-RGB theory, based on low-light blurred images The maximum value of each pixel in the three RGB channels is used to generate a luminance map. Brightness diagram The Middle Pixel value of each pixel As shown in the formula below:

[0072]

[0073] in, For low-light blurred images In the The first channel 1 pixel, It is a red channel. For green channel, The blue channel;

[0074] Based on brightness map Generate a saturation map and a preliminary chroma map, and the pixel value of any pixel in the saturation map. and the chromaticity components of any pixel in the preliminary chromaticity map As shown in the formula below:

[0075]

[0076]

[0077] in, Brightness diagram The Middle The pixel value of each pixel and the low-light blurred image The minimum value of pixels at the same position in the RGB channels The difference, For low-light blurred images Any pixel in the red channel The component values ​​in For low-light blurred images Any pixel in the green channel The component values ​​in For low-light blurred images Any pixel in the blue channel The component values ​​in the modulo operation;

[0078] Due to the red channel The component values ​​in the chromaticity component are at the hue endpoints. and There is a mathematical discontinuity at this point, which may produce red-fragmented noise after enhancing a low-light blurred image. To solve this problem, the chroma component... Mapped to two orthogonal components and To achieve polarization representation, orthogonal components and As shown in the formula below:

[0079]

[0080] Considering the problem of excessive amplification of color noise in low-light regions within the HSV color space, this embodiment introduces an adaptive light intensity compression function. This function maintains stable Euclidean distances between similar colors through nonlinear compression, thus constructing a first adaptive light intensity compression function. As shown in the formula below:

[0081]

[0082] in, To control the compression intensity in low-light regions using learnable parameters, Here is a constant used to prevent gradient explosion:

[0083] The first adaptive light intensity compression function Saturation diagram Separately and polarized orthogonal components , Element-wise multiplication yields two intermediate orthogonal components. and As shown in the formula below:

[0084]

[0085] The orthogonal components and By stitching along the channel dimension, a chroma map is obtained. ;

[0086] Low-light enhancement encoders such as Figure 2 As shown, it includes multiple cascaded cross-attention layers HVI-CA and multiple fusion layers HVI-Fuse;

[0087] Each cross-attention layer HVI-CA includes a multi-head cross-attention module, a luminance enhancement module, and a chrominance denoising module;

[0088] The multi-head cross-attention module consists of two branches: a chroma branch for processing the chroma map and a luma branch for processing the luma map. The chroma branch processes the chroma map... The input is a chromaticity branch, which aims to model color invariance features by using the continuity of orthogonal components to suppress noise in dark areas and correct chromaticity deviations, thereby maintaining realistic texture and color consistency and obtaining chromaticity features; the luminance branch uses a luminance map... Using the light distribution as input, this branch focuses on learning the physical constraints of the light distribution to obtain brightness features, thereby achieving adaptive brightness enhancement and global illuminance estimation, and avoiding overexposure and underexposure problems.

[0089] The chroma branch and the luminance branch achieve bidirectional feature interaction through a multi-head cross-attention mechanism: the chroma branch provides the luminance branch with prior information on noise distribution to guide the luminance enhancement intensity in low-light areas; at the same time, the luminance branch provides the chroma branch with illumination intensity weights so as to focus on noise reduction in dark areas and on detail and texture preservation in bright areas, thereby achieving joint optimization of global illumination and local color.

[0090] After passing through the multi-head cross-attention module, the chroma features output by the chroma branch and the luminance features output by the luminance branch are respectively entered into two independent functional modules for targeted processing: the chroma denoising module receives the chroma features, suppresses noise in dark areas and corrects chroma deviations, and outputs a denoised chroma feature map. The brightness enhancement module receives brightness features, learns the physical constraints of light distribution, prevents overexposure and underexposure, achieves adaptive brightness adjustment, and outputs an enhanced brightness feature map. ;

[0091] The chroma denoising module comprises multiple convolutional sub-modules. The chroma features input to the chroma denoising module are processed through the first convolutional sub-module, and then fed into the second and third convolutional sub-modules respectively. By dividing the chroma features into two parallel branches, complementary modeling of chroma information can be performed from different feature subspaces, enhancing the network's ability to represent complex noise patterns and subtle chroma variations. The outputs of the second and third convolutional sub-modules are non-linearly mapped using the Tanh activation function to enhance the expressive power of the chroma features and constrain their numerical range. These are then added to the outputs of the first convolutional sub-module to obtain two chroma enhancement features. An element-wise multiplication fusion operation is performed on the two chroma enhancement features to achieve mutual constraint and enhanced consistency between the two features. Subsequently, the fused features are input to the fourth convolutional sub-module for further modeling; the output of the fourth convolutional sub-module is added to the chroma features input to the chroma denoising module to obtain the denoised chroma feature map. ;

[0092] The brightness enhancement module comprises multiple convolutional sub-modules. The brightness features input to the brightness enhancement module are processed through the fifth convolutional sub-module, and then fed into the sixth and seventh convolutional sub-modules respectively. The outputs of the sixth and seventh convolutional sub-modules are non-linearly mapped using the Tanh activation function, and then their residuals are added to the output of the fifth convolutional module to obtain two intermediate brightness features. These two intermediate brightness features are then fused through element-wise multiplication and fed into the eighth convolutional sub-module. The output of the eighth convolutional sub-module is then added to the residuals of the input brightness features to obtain the enhanced brightness feature map. ;

[0093] The HVI-Fuse fusion layer comprises multiple convolutional modules and a multi-stage compressed activation fusion module (MSEF). The HVI-Fuse fusion layer is used to fuse denoised chroma feature maps. and enhanced brightness feature map The specific method is as follows:

[0094] Two convolutional modules are used to process the denoised chrominance feature map. and enhanced brightness feature map Perform convolution and utilize the enhanced brightness features resulting from the convolution. Denoising chromaticity features after convolution Linear weighted enhancement is performed to address the problem of chromaticity information attenuation under low light conditions, resulting in preliminary enhanced chromaticity features. As shown in the formula below:

[0095]

[0096] in, This is the luminance fusion coefficient, used to control the degree to which luminance information compensates for chromaticity features.

[0097] This will initially enhance color characteristics. The input is fed into the Multi-Stage Compressed Activation Fusion (MSEF) module, which simultaneously enhances salient features and suppresses background noise in both the spatial and channel domains to obtain the final enhanced chroma features. ,

[0098] The Multi-Stage Compressed Activation Fusion (MSEF) module includes layer normalization, global average pooling, ReLU activation function, Tanh activation function, and depthwise convolution, ultimately enhancing chroma features. As shown in the formula below:

[0099]

[0100] in, For global average pooling, , For trainable weight parameters, For depthwise convolution, This is a compressed channel feature map used to capture the dependencies between channels. The expanded attention activation map is used as a mask to weight the features. To ultimately enhance chromaticity features, It is the ReLU activation function. For Tanh activation function, For layer normalization;

[0101] This will ultimately enhance the chromaticity features. and the enhanced brightness feature map after convolution The features are concatenated by channel and passed through a convolutional module to obtain the fused and enhanced feature map output by the fusion layer HVI-Fuse;

[0102] This embodiment employs an encoder-decoder structure to uniformly handle low-light enhancement and deblurring tasks. The encoder incorporates eight cross-attention layers and four fusion layers, and the denoised chroma feature map output from the first cross-attention layer is used... and enhanced brightness feature map Input the second cross-attention layer and the first fusion layer respectively; denoise the chroma feature map output by the second cross-attention layer. and enhanced brightness feature map Input to the third cross-attention layer; denoise the chroma feature map output by the third cross-attention layer. and enhanced brightness feature map The inputs are respectively fed into the fourth cross-attention layer and the second fusion layer; the denoised chroma feature map output from the fourth cross-attention layer is then processed. and enhanced brightness feature map The input is fed into the fifth cross-attention layer, and the denoised chroma feature map output by the fifth cross-attention layer is used. and enhanced brightness feature map Input to the sixth cross-attention layer; denoise the chroma feature map output by the sixth cross-attention layer. and enhanced brightness feature map The inputs are fed into the seventh cross-attention layer and the third fusion layer, respectively; the denoised chroma feature map output from the seventh cross-attention layer is then processed. and enhanced brightness feature map Input to the eighth cross-attention layer; denoise the chroma feature map output by the eighth cross-attention layer. and enhanced brightness feature map Input to the fourth fusion layer;

[0103] By using the shared feature bridge connection of the HVI-Fuse fusion layer, the brightness, structure, and sharpness of the image are collaboratively optimized, avoiding the quality loss caused by the independent processing of low-light enhancement and deblurring in traditional methods. The dual-branch cross-attention mechanism achieves natural brightness enhancement and color balance of the image in low-light environments through the collaborative enhancement of brightness and chromaticity features, effectively avoiding artifacts and color cast problems common in traditional low-light enhancement methods.

[0104] The deblurring decoder includes multiple fuzzy-aware reconstruction layers MBR-WR, which are used to perform layer-by-layer deblurring and refinement reconstruction of the fused enhanced feature map to generate a preliminary reconstructed image. The fuzzy-aware reconstruction layer MBR-WR is used to generate the preliminary reconstructed image and includes a multi-scale fuzzy-aware residual module MBR and a wavelet reconstruction module WRM. Each multi-scale fuzzy-aware residual module MBR is connected to a wavelet reconstruction module WRM.

[0105] Multi-scale fuzzy perception residual module MBR, such as Figure 3 As shown, it includes multiple depthwise separable convolutional sub-modules, a multi-scale feature extraction branch (MFEB), and a fuzzy attention perception branch (BAPB). The multi-scale fuzzy perception residual module (MBR) takes the fused enhanced feature map output from the fusion layer (HVI-Fuse) and the feature map output from the previous fuzzy perception reconstruction layer (MBR-WR) as input, or only the feature map output from the previous fuzzy perception reconstruction layer (MBR-WR) as input. Through a dual-branch structure, it collaboratively extracts fuzzy features of different directions and scales in both spatial and channel dimensions, and performs targeted reconstruction. The process of the fuzzy perception reconstruction layer (MBR-WR) generating the initial reconstructed image is as follows:

[0106] The feature map of the input multi-scale fuzzy perception residual module MBR is processed using the first depthwise separable convolutional submodule. The first depthwise separable convolutional submodule is processed, comprising a 3×3 depthwise separable convolution DSConv kernel and a LeakyReLU activation function; the feature map output by the first depthwise separable convolutional submodule is processed. Input the multi-scale feature extraction branch MFEB and the fuzzy attention perception branch BAPB respectively;

[0107] The Multi-Scale Feature Extraction (MFEB) branch is used to extract structural detail information within different receptive fields, enhancing the model's ability to recover regions with varying degrees of blur. Specifically, the feature map input to the MFEB branch is... Simultaneously, multiple parallel depth-dilated convolutional submodules are input. In this embodiment, four parallel depth-dilated convolutional submodules are set, with dilation rates of 1, 2, 4, and 8 respectively. The kernel size of the depth-dilated convolution in each submodule is 3×3 to capture spatial fuzzy features at different scales. Subsequently, the output features of the multiple parallel depth-dilated convolutional submodules are concatenated along the channel dimension to obtain multi-scale fused features. To further integrate features at different scales and control channel dimensions, a depthwise separable dilated convolutional submodule with a dilation rate of 1 and a kernel size of 3×3 is used to fuse features at multiple scales. Perform compression mapping to obtain fused features. Next, the fusion features will be... The input is fed into the channel attention submodule SE, which extracts global information through global average pooling and adaptively calculates the weight coefficients of each channel through dimensionality reduction and expansion operations. This highlights key texture features and suppresses redundant responses, outputting enhanced features. Finally, the enhanced features will be... Feature maps of the input multi-scale feature extraction branch MFEB Residual connections are formed by element-wise addition to obtain multi-scale feature extraction results. The multi-scale feature extraction branch (MFEB) can effectively enhance the quality of cross-scale feature interaction and blurred edge reconstruction while maintaining feature consistency.

[0108] The Fuzzy Attention Perception Branch (BAPB) is used to detect and enhance directional and local blur features in motion-blurred regions. Its structure is based on a multi-kernel strip pooling mechanism, which processes the feature maps input to the BAPB branch. pass Strip pooling operations at different scales are performed to extract directional fuzzy information along the horizontal and vertical directions, respectively, to obtain... The strip features are divided into strips at different scales, including horizontal and vertical features at each scale. In this embodiment, the scales of the strip pooling operation are set to 1, 3, 5, and 7. The horizontal and vertical features at each scale are fused by 1D or 2D convolution to form fuzzy response tensors at different scales. , Then, the fuzzy response tensors of all scales are concatenated and input into two 1×1 convolutional submodules for processing. The features output by the second 1×1 convolutional module are then combined with the feature maps input to the fuzzy attention perception branch BAPB. Element-level weighted fusion is performed to obtain intermediate features The first 1×1 convolutional submodule includes a 1×1 convolution and a ReLU activation function, and the second 1×1 convolutional submodule includes a 1×1 convolution and a Sigmoid activation function; intermediate features A blurred attention mask is generated by mapping two layers of 3×3 convolutions with a sigmoid activation function. This is used to adaptively adjust the response intensity of blurred regions globally, thereby enhancing features at significantly blurred edges and suppressing noise in smooth regions. Ultimately, intermediate features... With fuzzy attention mask Element-level weighted fusion is performed to obtain fuzzy enhanced features. .

[0109] To integrate the structural and directional information extracted by the Multi-Scale Feature Extraction (MFEB) branch and the Fuzzy Attention Perception (BAPB) branch, the features output by the multi-scale feature extraction branch are... Features of the output of the fuzzy attention perception branch The splicing is performed along the channel dimension to obtain the splicing features. Subsequently, the splicing features will be... The input is fed into a second-depth separable convolutional submodule, where feature integration and channel compression are performed to obtain intermediate fused features. To ensure the continuity of feature propagation and accelerate network convergence, intermediate fused features are... Feature maps of the input multi-scale fuzzy perception residual module (MBR) Establish cross-branch residual connections to obtain intermediate enhanced feature maps from the output of the multi-scale fuzzy perception residual module (MBR). Through the aforementioned parallel dual-branch and residual fusion structure, adaptive joint modeling of global fuzzy direction and local fuzzy degree can be achieved, providing higher quality feature input for subsequent enhancement.

[0110] Wavelet reconstruction module, such as Figure 4 As shown, it includes:

[0111] The intermediate enhanced feature map output by the multi-scale fuzzy perception residual module MBR The input is fed into the wavelet reconstruction module WRM, which includes the discrete wavelet transform unit DWT and the discrete wavelet inverse transform unit IDWT.

[0112] Discrete Wavelet Transform (DWT) unit utilizes Haar wavelets to enhance intermediate feature maps. Decomposed into the wavelet frequency domain, four sub-band feature maps of different frequencies are generated. Wavelet transform yields the low-frequency and high-frequency component information, specifically including:

[0113]

[0114] in, This refers to the low-frequency sub-band, which contains information about overall brightness and structure. , and Represents high-frequency subbands containing texture and edge detail information in different directions. For horizontal high-frequency sub-band, For vertical high-frequency sub-band, For diagonal high-frequency sub-bands, This is a discrete wavelet transform unit. Discrete wavelet transform can effectively separate structural and detailed features, reducing interference when directly recovering high-frequency textures in the spatial domain.

[0115] To suppress coupling interference between different frequency sub-bands and to specifically recover high-frequency details in blurred regions, each sub-band is input into a feature recovery network for independent reconstruction, resulting in the recovered features of each sub-band:

[0116]

[0117] in, This represents a feature recovery network consisting of three convolutional layers, with each convolutional layer followed by a non-linear activation function to enhance feature representation. After this process, To reconstruct the low-frequency subband, used to reconstruct the overall illumination structure; To reconstruct the horizontal high-frequency subband, To reconstruct the vertical high-frequency subband, To reconstruct the diagonal high-frequency subband, the reconstructed high-frequency subband is used to restore details, edges, and textured areas, thereby enhancing blurred and low-light areas with multi-directional high-frequency enhancement.

[0118] Finally, the low-frequency subband will be reconstructed. Reconstructing horizontal high-frequency subband Reconstructing vertical high-frequency subbands and reconstructing diagonal high-frequency subband The image is input into the Discrete Wavelet Inverse Transform (IDWT) unit for wavelet-domain to spatial domain reconstruction, yielding a preliminary reconstructed image. As shown in the formula below:

[0119]

[0120] Through this inverse transformation process, different frequency components are re-fused in the spatial domain, achieving the joint restoration of the image's structural consistency and sense of detail.

[0121] The multi-scale blur perception residual module can extract global and local blur features at different scales, accurately recovering various blurred regions of the image and avoiding the neglect of blurred regions and over-sharpening. By performing frequency domain separation processing through wavelet transform, low-frequency structural information and high-frequency detail texture are recovered separately, further improving the image's detail recovery capability, especially maintaining clear edges and textures under complex backgrounds and blurry conditions.

[0122] The initial reconstructed image output by the deblurring decoder. The input convolutional mapping layer and the inverse HVI transform module PHVIT generate a three-channel feature representation based on the HVI color space. Then, the inverse HVI transform module PHVIT maps the three-channel feature representation to the RGB color space to obtain the enhanced instrument image.

[0123] The Inverse HVI Transform (PHVIT) module is used to achieve stable reconstruction from the HVI color space to the RGB color space. PHVIT takes the three-channel feature representation output from the convolutional mapping layer as input and decomposes it into two orthogonal chromaticity components. , and brightness component Based on luminance components Constructing a second adaptive light intensity compression function As shown in the formula below:

[0124]

[0125] in, These are learnable parameters used to control the compression intensity in low-light regions. This is a constant used to prevent gradient explosion. Luminance component The Middle Pixel values ​​of each pixel:

[0126] Using the second adaptive light intensity compression function For chromaticity components and Normalization scaling is performed to effectively suppress noise amplification and color oversaturation during brightness enhancement, resulting in corrected chromaticity components. and As shown in the formula below:

[0127]

[0128] Based on this, the corrected chromaticity components , and brightness component Mapping to the HSV color space is shown in the following formula:

[0129]

[0130] in, and Both are learnable linear adjustment parameters, used to adaptively control the saturation and brightness amplitude of the image to improve the visual consistency of the enhancement results under different lighting conditions.

[0131] Based on the standard HSV to RGB segmented mapping rules, the obtained HSV color space representation is converted to sRGB color space, and the final enhanced instrument image is output.

[0132] A CLIP-based semantic guidance-guided cue contrast learning mechanism is established to train the image enhancement network;

[0133] In this embodiment, to further improve the enhancement performance and semantic consistency of the image enhancement network under low-light blurred image conditions, a CLIP-based semantically guided cue contrast learning mechanism is introduced. By constructing multimodal semantic constraints, a balance between visual sharpness and semantic fidelity in the enhancement results is achieved. Specifically, the CLIP-based semantically guided cue contrast learning mechanism includes a learnable cue pair construction phase and an enhancement network semantic guidance phase.

[0134] In the learnable cue pair construction phase, firstly, a pair of learnable text cue vectors is constructed to represent different image quality states. The aim is to accurately anchor the "sharp" and "degraded" features in the semantic space through contrastive learning. The learnable text cue vectors include positive cue vectors. and negative tips ,in This represents the total number of tokens for learnable text prompt vectors. Feature dimensions for each token. To give the prompts explicit semantic priors, positive prompts... Initialize the semantic embedding corresponding to the natural language description "a clear and bright instrument image", and include negative prompts. Initialize it to describe the semantic embedding corresponding to "a low-light blurred instrument image".

[0135] Subsequently, the parameters frozen in the CLIP model were used to correct the prompt. and negative tips Perform iterative optimization. Reduce low-light blurred images. Clear image with reference light The image encoders of the CLIP model are input respectively. This is done to extract latent image semantic features; simultaneously, positive prompts will be displayed. and negative tips Text encoder for input CLIP Extract semantic embedding representations. Define a cue contrast loss function under different lighting and blur conditions by calculating the semantic space similarity between the image and text:

[0136]

[0137] in, For degradation category labels, Indicates a low-light blurred image. Indicates a clear, bright image; For predicting images The probability of belonging to the positive suggestion category is calculated using the Softmax-normalized cosine similarity:

[0138]

[0139] Minimize the comparison loss function by hinting. Only for positive prompts and negative tips Perform gradient updates to make positive suggestions and negative tips In the semantic space, it can accurately distinguish between degenerate and clear scenes, thereby obtaining optimized semantic priors with strong discriminative power.

[0140] After optimizing the suggestion vector, the positive suggestion will be... and negative tips Fixed and embedded into the training phase of the image augmentation network. An augmentation loss based on CLIP is introduced. Used to constrain low-light blurred images Enhanced image Approximate positive prompt semantics in semantic space CLIP-based augmentation loss As shown in the formula below:

[0141]

[0142] By minimizing CLIP-based enhancement loss This forces low-light blurring of images Enhanced image Semantically, it avoids negative cues that represent ambiguity and darkness. And to the positive reminder representing clarity and light. Convergence. This multimodal consistency training strategy enables the image enhancement network to achieve joint enhancement learning of low-light blurred images at both the semantic and pixel levels, effectively improving the model's adaptability to the dual degradation conditions of insufficient lighting and motion blur. This results in output images that are significantly superior to traditional low-light or deblurring single-task models in terms of brightness, texture detail, and semantic consistency.

[0143] A low-light blurred instrument image is acquired, and the low-light blurred image to be enhanced is processed based on a trained image enhancement network to obtain the corresponding enhanced image.

[0144] Example 2:

[0145] This embodiment proposes an electronic device, including: one or more processors, and a memory for storing instructions, which, when executed by the one or more processors, cause the one or more processors to perform the combined low-light enhancement and deblurring instrument image enhancement method.

[0146] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program that, when executed by the processor, implements the instrument image enhancement method of joint low-light enhancement and deblurring as described in the embodiments. It is understood that the electronic device may also include an input / output (I / O) interface and communication components.

[0147] The processor is used to execute all or part of the steps in the instrument image enhancement method with combined low-light enhancement and deblurring as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.

[0148] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the instrument image enhancement method with combined low-light enhancement and deblurring described in the above embodiments.

[0149] Example 3:

[0150] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0151] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the instrument image enhancement method with combined low-light enhancement and deblurring described in the various embodiments of this application.

[0152] The aforementioned storage media include: flash memory, hard disk, multimedia card, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disk, optical disk, server, APP (Application) application store, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned combined low-light enhancement and deblurring instrument image enhancement method.

[0153] Example 4:

[0154] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned method for enhancing instrument images by combining low-light enhancement and deblurring.

[0155] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0156] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0157] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of this disclosure and its equivalents, then the intent of this disclosure also includes these modifications and variations.

Claims

1. A method for enhancing instrument images by combining low-light enhancement and deblurring, characterized in that, Includes the following steps: An image enhancement network for instrument image enhancement is constructed to perform low-light enhancement and deblurring on low-light blurred instrument images. The network includes a decomposition module, a low-light enhancement encoder, multiple fusion layers, a deblurring decoder, a convolutional mapping layer, and an inverse HVI transform module. A CLIP-based semantic guidance-guided prompting contrast learning mechanism is established to train the image enhancement network; Obtain the low-light blurred instrument image to be enhanced, and perform low-light enhancement and deblurring processing on the low-light blurred image to be enhanced based on the trained image enhancement network to generate the enhanced instrument image; The decomposition module is used to decompose the input low-light blurred image into chroma and luminance, and decompose the low-light blurred instrument image into a luminance map and a chroma map. The low-light enhancement encoder is used to jointly enhance the luminance and chrominance of the luminance map and chrominance map to generate a denoised chrominance feature map and an enhanced luminance feature map. The fusion layer is used to fuse the denoised chromaticity feature map and the enhanced luminance feature map into a fused enhanced feature map; the deblurring decoder is used to perform layer-by-layer deblurring and thinning reconstruction on the fused enhanced feature map to generate a preliminary reconstructed image; the convolutional mapping layer and the inverse HVI transform module are used to perform color space conversion on the preliminary reconstructed image to obtain the enhanced instrument image.

2. The method for enhancing instrument images by combining low-light enhancement and deblurring according to claim 1, characterized in that, The specific method for decomposing a low-light blurred instrument image into a luminance map and a chrominance map is as follows: A luminance map is generated based on the maximum value of each pixel in the three RGB channels of the low-light blurred image. A saturation map and a preliminary chromaticity map are generated based on the luminance map. The chromaticity component of each pixel in the preliminary chromaticity map is mapped to two polarized orthogonal components. A first adaptive light intensity compression function is established based on the brightness map. The first adaptive light intensity compression function and the saturation map are multiplied element-wise by the two polarized orthogonal components to obtain two intermediate orthogonal components. By splicing the two intermediate orthogonal components along the channel dimension, a chromaticity map is obtained.

3. The method for enhancing instrument images by combining low-light enhancement and deblurring according to claim 1, characterized in that, The low-light enhancement encoder includes multiple cross-attention layers; The specific method for joint enhancement of luminance and chrominance in low-light enhancement encoders is as follows: The luminance and chrominance maps are input into the low-light enhancement encoder and passed through multiple cross-attention layers to generate a luminance-equalized image. Each cross-attention layer includes a multi-head cross-attention module, a brightness enhancement module, and a chroma denoising module; The multi-head cross-attention module includes a chroma branch and a luminance branch, which are used to process the chroma map and luminance map respectively, extract chroma features and luminance features, input the chroma features into the chroma denoising module, and input the luminance features into the luminance enhancement module. The chroma denoising module includes multiple convolutional sub-modules. The chroma features input to the chroma denoising module are processed by the first convolutional sub-module and then input into the second and third convolutional sub-modules respectively. The outputs of the second and third convolutional sub-modules are non-linearly mapped by the Tanh activation function and then added to the output of the first convolutional module by residual addition to obtain two chroma enhancement features. An element-wise multiplication fusion operation is performed on the two chroma enhancement features. The fused features are then input into the fourth convolutional submodule. The output of the fourth convolutional submodule is added to the residual of the chroma features input to the chroma denoising module to obtain the denoised chroma feature map. The brightness enhancement module includes multiple convolutional sub-modules. The brightness features input to the brightness enhancement module are processed through the fifth convolutional sub-module and then input into the sixth and seventh convolutional sub-modules, respectively. The outputs of the sixth and seventh convolutional sub-modules are nonlinearly mapped using the Tanh activation function, and then the residuals are added to the output of the fifth convolutional module to obtain two intermediate brightness features. After performing an element-wise multiplication fusion operation on the two intermediate brightness features, the fused features are input into the eighth convolutional sub-module. The output of the eighth convolutional sub-module is added to the residuals of the brightness features input to the brightness enhancement module to obtain the enhanced brightness feature map.

4. The method for enhancing instrument images by combining low-light enhancement and deblurring according to claim 1, characterized in that, The fusion layer consists of multiple convolutional modules and a multi-stage compressed activation fusion module; The specific method for fusing the denoised chromaticity feature map and the enhanced luminance feature map into a fused enhanced feature map is as follows: Two convolutional modules are used to convolve the denoised chroma feature map and the enhanced luminance feature map respectively. The enhanced luminance feature after convolution is used to linearly weight the denoised chroma feature after convolution to obtain the preliminary enhanced chroma feature. The initial enhanced chroma features are input into the multi-stage compression activation fusion module to obtain the final enhanced chroma features; The final enhanced chroma feature and the convolutional enhanced luminance feature map are concatenated by channel and passed through a convolution module to obtain the fused enhanced feature map.

5. The method for enhancing instrument images by combining low-light enhancement and deblurring according to claim 1, characterized in that, The deblurring decoder includes multiple fuzzy perception reconstruction layers; each fuzzy perception reconstruction layer includes a multi-scale fuzzy perception residual module and a wavelet reconstruction module, with each multi-scale fuzzy perception residual module connected to a wavelet reconstruction module; the multi-scale fuzzy perception residual module includes multiple depthwise separable convolutional sub-modules, a multi-scale feature extraction branch, and a fuzzy attention perception branch; The specific method for layer-by-layer deblurring and thinning reconstruction of the fused and enhanced feature map is as follows: Multiple fused and enhanced feature maps output from multiple fusion layers are input into the deblurring decoder, and after passing through multiple blur-aware reconstruction layers, a preliminary reconstructed image is generated, including: The fused enhanced feature map and the feature map output from the previous fuzzy perception reconstruction layer are used as input, or only the feature map output from the previous fuzzy perception reconstruction layer is used as input, and then passed through the multi-scale fuzzy perception residual module and the wavelet reconstruction module in sequence. In the multi-scale fuzzy perception residual module, a first depthwise separable convolutional submodule is used to process the feature map input to the multi-scale fuzzy perception residual module. The first depthwise separable convolutional submodule is processed, comprising a 3×3 depthwise separable convolution DSConv kernel and a LeakyReLU activation function; the feature map output by the first depthwise separable convolutional submodule is processed. Input the multi-scale feature extraction branch and the fuzzy attention perception branch respectively; Features output from the multi-scale feature extraction branch Fuzzy Enhancement Features of Fuzzy Attention Perception Branch Output The splicing is performed along the channel dimension to obtain the splicing features. ; splicing features The input is fed into a second-depth separable convolutional submodule, where feature integration and channel compression are performed to obtain intermediate fused features. ; integrate intermediate features Feature maps of the input multi-scale fuzzy perception residual module Establish cross-branch residual connections to obtain intermediate enhanced feature maps from the output of the multi-scale fuzzy perception residual module (MBR). ; intermediate enhanced feature map The input is fed into the wavelet reconstruction module to generate a preliminary reconstructed image. .

6. The method for enhancing instrument images by combining low-light enhancement and deblurring according to claim 5, characterized in that, The multi-scale feature extraction branch includes multiple parallel depth-wise dilated convolution sub-modules, depth-separable dilated convolution sub-modules, and channel attention sub-modules; Multiple parallel depthwise dilated convolutional submodules are used to extract feature maps from the input multi-scale feature branches. Convolution is performed, and the output features of multiple parallel depthwise dilated convolution sub-modules are concatenated along the channel dimension to obtain multi-scale fused features. ; Multi-scale feature fusion using depthwise separable dilated convolution submodules Perform compression mapping to obtain fused features. ; Fusion features Input is fed into the channel attention submodule, and the output is attention-enhanced features. ; Attention-enhancing features Feature maps of the input multi-scale feature extraction branch The residual connections are formed by element-wise addition, resulting in the features output by the multi-scale feature extraction branch. .

7. The method for enhancing instrument images by combining low-light enhancement and deblurring according to claim 5, characterized in that, The fuzzy attention perception branch includes multiple strip pooling operations of different scales, two 1×1 convolutional sub-modules, and two 3×3 convolutional sub-modules; the first 1×1 convolutional sub-module includes a 1×1 convolution and a ReLU activation function, and the second 1×1 convolutional sub-module includes a 1×1 convolution and a Sigmoid activation function; the first 3×3 convolutional sub-module includes a 3×3 convolution and a ReLU activation function, and the second 3×3 convolutional sub-module includes a 3×3 convolution and a Sigmoid activation function; Strip pooling operations of different scales were used to refine the feature maps of the fuzzy attention perception branch along the horizontal and vertical directions, respectively. Directional fuzzy information is extracted to obtain fuzzy response tensors at different scales; The fuzzy response tensors at all scales are concatenated and input into two 1×1 convolutional submodules for processing. The features output from the second 1×1 convolutional module are then combined with the feature maps input to the fuzzy attention perception branch. Element-level weighted fusion is performed to obtain intermediate features ; Use two 3×3 convolutional submodules for intermediate features The fuzzy attention mask output by the second 3×3 convolutional submodule is processed. and intermediate features Element-level weighted fusion is performed to obtain fuzzy enhanced features. .

8. The method for enhancing instrument images by combining low-light enhancement and deblurring according to claim 5, characterized in that, The wavelet reconstruction module includes a discrete wavelet transform unit and a discrete wavelet inverse transform unit; Generate final enhanced output features The specific method is as follows: intermediate enhanced feature map The input is fed into the wavelet reconstruction module, which uses Haar wavelets to enhance the intermediate feature maps. Decomposed into the wavelet frequency domain, sub-band feature maps of different frequencies are generated, including low-frequency sub-band feature maps, horizontal high-frequency sub-band feature maps, vertical high-frequency sub-band feature maps, and diagonal high-frequency sub-band feature maps; Subband feature maps of different frequencies are input into the feature recovery network for independent reconstruction, resulting in multiple reconstructed subband feature maps, including reconstructed low-frequency subband feature maps, reconstructed horizontal high-frequency subband feature maps, reconstructed vertical high-frequency subband feature maps, and reconstructed diagonal high-frequency subband feature maps. Multiple reconstructed sub-band feature maps are input into the discrete wavelet inverse transform unit to obtain a preliminary reconstructed image. .

9. The method for enhancing instrument images by combining low-light enhancement and deblurring according to claim 1, characterized in that, The specific method for color space conversion of the initially reconstructed image is as follows: The initially reconstructed image is sequentially input into a convolutional mapping layer and an inverse HVI transform module. The convolutional mapping layer outputs a three-channel feature representation of the initially reconstructed image, and the inverse HVI transform module decomposes the three-channel feature representation into two orthogonal chromaticity components. , And a brightness component; A second adaptive light intensity compression function is constructed based on the luminance component, and this function is then used to compress the two orthogonal chrominance components. and Normalization and scaling are performed to obtain the corrected chromaticity components. and ; Corrected chromaticity components , The luminance component is mapped to the HSV color space. Based on the standard HSV to RGB segmented mapping rules, the resulting HSV color space representation is converted to the sRGB color space, and the enhanced instrument image is output.

10. The method for enhancing instrument images by combining low-light enhancement and deblurring according to claim 1, characterized in that, The specific method for training the image enhancement network by establishing a CLIP-based semantic prompting contrast learning mechanism is as follows: Construct a pair of learnable text cue vectors to represent different image quality states. The learnable text cue vectors include positive cue and negative cue. The positive cue is initialized with a semantic embedding describing a clear instrument image in good light, and the negative cue is initialized with a semantic embedding describing a blurry instrument image in low light. The CLIP model is used to iteratively optimize positive and negative cues. The low-light blurred image and the reference bright image are respectively input into the CLIP model's image encoder to extract latent semantic features. The text encoder that inputs positive and negative prompts into CLIP extracts semantic embedding representations; We construct a cue contrast loss function under illumination and blur conditions, and update the gradients of positive and negative cues by minimizing the cue contrast loss function; The iteratively optimized positive and negative cues are fixed and embedded into the training phase of the image augmentation network. A CLIP-based augmentation loss is established for training the image augmentation network. The trained image augmentation network is obtained by minimizing the CLIP-based augmentation loss.

Images