Image enhancement method based on priori guided wavelet state space model

Abstract

The application discloses an image enhancement method based on a priori guided wavelet state space model, comprising the following steps: S1, acquiring an original image, converting the original image to a YCbCr color space, and extracting a luminance component and a chrominance component; S2, extracting a basic prior mask probability graph through a convolutional neural network, and performing step-by-step down-sampling to generate a multi-scale prior pyramid; S3, inputting the luminance component into a main enhancement branch to perform feature extraction and enhancement, and obtaining an enhanced luminance tensor; and S4, splicing the enhanced luminance tensor and the original chrominance component along a channel dimension, and obtaining a final enhanced image through color space inversion. The application solves the defects that a traditional enhancement network is prone to amplifying mine "halo" and "dust fog", smoothes the unique high-frequency particle noise of the mine, greatly improves the definition and signal-to-noise ratio of the image, and avoids irreversible loss of detail information.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and digital image processing technology, and in particular relates to an image enhancement method based on the prior guided wavelet state space model (Mamba) suitable for complex environments such as mines or tunnels. Background Technology

[0002] In enclosed underground environments such as mines and tunnels, the lack of natural light and explosion-proof requirements often result in extremely low illumination in the acquired visual images. Simultaneously, construction or mining operations generate large amounts of suspended dust, leading to severe fogging and high-frequency particle noise in the captured images. High-quality mine images are crucial for underground safety monitoring, automated equipment operation, and disaster early warning. Therefore, developing an effective image enhancement method is essential.

[0003] In general, methods for enhancing degraded images can be divided into three categories: image enhancement methods based on physical models; image enhancement methods based on non-physical models; and image enhancement methods based on deep learning. Among these, image enhancement methods based on physical models first establish a mathematical model based on the imaging mechanism of degraded images in complex mines, then analyze the unknown parameters in the model using manual prior knowledge or assumptions, and inversely derive the enhanced, clear image. This type of method can achieve certain enhancement effects in specific degraded scenarios, but due to the harshness and variability of the complex underground mine environment, the prior knowledge is prone to instability, resulting in varying robustness and stability of the enhancement effect.

[0004] Image enhancement methods based on non-physical models often improve overall or local contrast and color accuracy by modifying pixel values ​​or histogram distribution. While these methods have low computational complexity, they lack consideration of the degradation mechanisms of degraded images, resulting in insufficient detail in the enhancement results.

[0005] Deep learning-based image enhancement methods leverage the powerful feature learning and computational capabilities of deep learning networks to obtain the final enhanced image. These methods can improve the visual quality of degraded images to some extent, but they typically require training the deep learning network on a large dataset containing pairs of "degraded images - clear images." However, such datasets are difficult to obtain under complex and harsh weather conditions, leading to significant limitations in these methods.

[0006] Current low-light image enhancement techniques primarily focus on restoring image brightness and color. In recent years, networks based on state-space models (such as the Mamba framework) have demonstrated excellent performance in long-range dependency modeling. However, existing techniques have significant drawbacks when applied to mining environments: conventional state-space models or wavelet enhancement networks, while improving global brightness, mistakenly amplify dust and particle noise in the image as effective high-frequency textures or global features. This results in severe "halo" and "dust" phenomena in the enhanced image, and the particle noise is sharpened, significantly reducing the visual quality of the image and the accuracy of subsequent machine vision tasks.

[0007] Existing technologies also include methods for decoupling noise in specific color spaces (such as YCbCr), but these methods typically rely on simple convolutional neural networks and lack the ability to model long-range global illumination. How to integrate noise decoupling mechanisms with long-range global sequence modeling capabilities at a mechanism-level within the network to achieve "enhancing detailed global illumination while suppressing dust noise" is a pressing technical problem in this field.

[0008] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0009] The purpose of this invention is to address the technical problem in existing technologies that easily amplify dust and high-frequency noise when performing low-light enhancement in a mining environment, and to provide an image enhancement method based on a priori guided wavelet state-space model.

[0010] To achieve the above objectives, the technical solution of the present invention is as follows: The image enhancement method based on the prior-guided wavelet state-space model includes the following steps: S1. Obtain the raw image It was then converted to the YCbCr color space, and the luminance component of the Y channel was extracted. chromaticity components of the CbCr channel ; S2. Using chromaticity components As input, a basic prior mask probability map is extracted using a convolutional neural network. And the basic prior mask probability map Perform stepwise downsampling to generate a multi-scale prior pyramid. Where i represents the scale; S3. Convert the brightness component The input is fed into the main enhancement branch for feature extraction and enhancement. The main enhancement branch adopts an encoder-decoder structure, and each layer of encoding includes the following steps: S3C1. Perform a single feature downsampling using a two-dimensional discrete wavelet transform on the input features, decomposing the input into low-frequency components. and high frequency components { , , The input features of the first layer are the luminance components. The input features of each subsequent layer are the output features of the previous layer. S3C2. Processing downsampled low-frequency components using a priori-guided low-frequency state-space model network PG-SSM. Spatial fusion characteristics were obtained. and spatial integration features with low-frequency components Pixel-level summation yields the deep coding features of this layer. The deep coding features of this layer The input is passed down to the next network layer until the final layer of the encoder, resulting in the final deep encoded features. Before scanning the state-space model network, the noise prior feature maps of the corresponding scale in the multi-scale prior pyramid are used as gating factors to perform element-wise attention weighting on the input features, thereby suppressing the feature intensity of high-noise regions. Each layer of decoding includes the following steps: S3D1. Decoder receives low-frequency characteristics of the input. The low-frequency features of this input are encoded as deep features in the final layer of the encoder in the first layer. The input features of each subsequent decoder layer are the output features of the previous decoder layer. At the same time, high-frequency components at the corresponding scale are extracted. The features are then spliced ​​and fused along the channel dimension to construct the original high-frequency features. , S3D2. Original high-frequency features Feature enhancement is performed to obtain updated high-frequency features. The formula is as follows:

[0011] in, It is a cross-band characteristic transfer function. Represents element-wise multiplication of spatial and channel dimensions. For learnable parameters, S3D3. Low-frequency features With updated high-frequency features The data is stitched together along the channel dimension and then subjected to a two-dimensional inverse wavelet transform. Perform feature upsampling to obtain the decoded features of the current level. Decoding features at this level This data serves as input to the next level of the decoder for subsequent decoding. The decoder outputs the final decoded features, which are then compared with the initial luminance components. Pixel-level addition is performed to obtain the enhanced luminance tensor. , S4. Enhance the brightness tensor With primary chromaticity components The images are stitched together along the channel dimension, and the final enhanced image is obtained through inverse color space transformation. . Furthermore, step S3C2 includes the following steps: P1. Extracting the noise prior feature maps of the corresponding scales from the multi-scale prior pyramid. For low-frequency components Adaptive soft-gating is performed to obtain the gated low-frequency component. The formula is as follows:

[0012] in, For linear mapping operations, For element-wise multiplication of spatial and channel dimensions. For learnable parameters, P2. Gating low-frequency components Along different directions of the image Each part is flattened to generate a one-dimensional sequence. And through the discrete state-space model, through implicit state Input Mapped to output The formula is as follows:

[0013]

[0014] in, Represents implicit state. This represents the input one-dimensional feature sequence. Indicates the output. Here is the state transition matrix. and Let be the input projection matrix and the output projection matrix, respectively, and t represent the sequence index of the discretized one-dimensional feature. P3. Perform an inverse flattening operation on the output sequence in each direction according to the original scanning direction to restore the original two-dimensional spatial resolution, and then sum the channels to obtain the spatial fusion features. , P4. Spatial fusion features are achieved through jump residual connections. with low-frequency components Pixel-level summation yields the deep coding features of this layer. The deep coding features obtained As input to the next network layer, the information continues to be passed down until the final layer of the encoder, resulting in the final deep encoded features.

[0015] Furthermore, step S3D3 obtains the decoding features of the current level. Then, calculate the smooth decoding features of the current level. The formula is as follows: In the formula This represents the anti-aliased smooth convolution computation and uses the smooth decoding features of this level. It serves as the input to the next level of decoder for subsequent decoding.

[0016] Furthermore, It is set to 0 during network initialization.

[0017] The present invention also provides an image enhancement system based on a priori guided wavelet state space model, comprising: a color space conversion module, a multi-scale prior pyramid generation module, a main enhancement branch, and a color space inverse conversion module. The main enhancement branch is an encoder-decoder structure, each encoder layer includes a downsampling module and a priori guided low-frequency state space model network module, and each decoder layer includes a priori guided high-frequency enhancement module and an upsampling module.

[0018] In addition, the present invention also provides an image enhancement device based on a priori guided wavelet state space model, comprising: a memory, a processor, and an image enhancement program based on a priori guided wavelet state space model stored in the memory and executable on the processor, wherein the image enhancement program based on a priori guided wavelet state space model is configured to implement an image enhancement method based on a priori guided wavelet state space model.

[0019] The present invention also provides a storage medium storing an image enhancement program based on a priori guided wavelet state space model, wherein the image enhancement program based on the priori guided wavelet state space model, when executed, implements an image enhancement method based on the priori guided wavelet state space model.

[0020] By adopting the above technical solution, the present invention has the following beneficial technical effects: (1) This invention breaks through the limitations of the traditional simple chaining of noise reduction and enhancement modules, injecting noise priors extracted from the chromaticity space into the core state transition process of the state space model (Mamba). By introducing noise priors as a gating mechanism in the low-frequency state space module, the problem of Mamba mistaking dense dust areas for "global illumination features" and spreading them to the surroundings is effectively solved, fundamentally addressing the defect of traditional enhancement networks that easily amplify mine "halos" and "dust fog". At the same time, the zero-initialization design of learnable gating parameters ensures smooth gradients in the early stages of training, avoiding the "identity mapping trap" of the network outputting a completely black image.

[0021] (2) In the high-frequency detail correction stage, this invention introduces an adaptive soft-gated filtering and cross-band feature fusion mechanism. Noise priors are explicitly used as attention weights to adaptively multiply attenuate the dust and noise responses in the high-frequency features, and then low-frequency illumination and structural information aligned through linear mapping is incorporated. This mechanism sharpens the edges of structures such as rock walls and tracks while smoothing the high-frequency particle noise unique to mines, significantly improving image clarity and signal-to-noise ratio.

[0022] (3) By replacing the pooling and interpolation layers in the traditional convolutional neural network with discrete wavelet transform (DWT) and inverse wavelet transform (IWT), the complete frequency domain information is preserved in the process of changing the feature resolution, and the irreversible loss of detailed information is avoided.

[0023] (4) The multi-scale prior pyramid feature map extracted by the present invention can be directly visualized, clearly indicating the network's discrimination results for the "dust / noise" region and the "track / wall / texture" region, providing a reliable interpretable basis for visual algorithms in complex mining environments. Attached Figure Description

[0024] Figure 1 This is a flowchart of the present invention.

[0025] Figure 2 This is a system structure diagram of the present invention. Detailed Implementation

[0026] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0027] Example 1: As Figure 1 As shown, the present invention provides an image enhancement method based on a priori guided wavelet state-space model, comprising the following steps: S1. Obtain native RGB low-light images from the mine. It was then converted to the YCbCr color space, and the luminance component of the Y channel was extracted. chromaticity components of the CbCr channel .

[0028] In this space, the main structural, textural, and lighting information is concentrated in the luminance component, while most color shifts and high-frequency dust noise information are concentrated in the chromaticity component.

[0029] S2. Using chromaticity components As input, a basic prior mask probability map representing the spatial distribution of high-frequency noise and dust is extracted using a lightweight convolutional neural network. And the basic prior mask probability map Perform stepwise downsampling to generate a multi-scale prior pyramid. Where i represents the scale, this embodiment is divided into three scales, such as [ This multi-scale a priori pyramid It can be used as a noise prior feature map for the main enhancement branch.

[0030] The lightweight convolutional neural network described herein is a conventional technique. This invention provides a specific structure, as shown in the following formula:

[0031] in, This represents a convolution operation with a kernel size of 3x3, a stride of 1, and padding of 1. This represents the LeakyReLU activation function. This represents the Sigmoid activation function. The final output is the base prior mask probability map. The pixel probability values ​​are normalized to the [0, 1] interval.

[0032] To align with feature maps of different depths from the main augmentation network, the base prior mask probability map is... A multi-scale prior pyramid is generated by stepwise downsampling using two-dimensional discrete wavelet transform. The pixel values ​​in this pyramid represent the probability distribution of whether a corresponding region in the image space belongs to dust or noise.

[0033] S3. Convert the brightness component The input is fed into the main enhancement branch for feature extraction and enhancement. Preferably, the main enhancement branch can adopt the encoder-decoder structure of U-Net, including N layers of encoders and decoders; in this embodiment, N=3. It is divided into encoding step S3C and decoding step S3D. Each encoding layer includes the following steps: S3C1. In the nth layer, the input features are downsampled once. Specifically, a two-dimensional discrete wavelet transform can be used for feature downsampling, decomposing the input into four frequency band sub-images—low-frequency components. High-frequency components { , , }, as shown in the following formula:

[0034] in, This represents the discrete wavelet transform, where i represents the downsampling scale, and i also corresponds to the encoder level, such as the scale. Corresponding to the first layer, scale Corresponding to the second layer, scale This corresponds to layer 3; the input features of layer 1 are the luminance components. The input features of each subsequent layer are the output features of the previous layer. Note that this step only downsamples the input once, but because the input features of each subsequent layer are the output features of the previous layer, the scale will change as the number of layers increases.

[0035] S3C2. Processing downsampled low-frequency components using a priori-guided low-frequency state-space model network PG-SSM. Before scanning the state space, the noise prior feature maps of the corresponding scale in the multi-scale prior pyramid are used as gating factors to perform element-wise attention weighting on the input features, thereby suppressing the feature intensity of high-noise regions.

[0036] Specifically, this includes: P1. To prevent the prior-guided low-frequency state-space model network PG-SSM from erroneously propagating local dust features as global illumination features during global sequence scanning, a soft prior gating mechanism is introduced before feature serialization enters the state-space scanning. Noise prior feature maps at the corresponding scale are extracted. For low-frequency components Adaptive soft-gating is performed to obtain the gated low-frequency component. The formula is as follows:

[0037] in, For linear mapping operations, For element-wise multiplication of spatial and channel dimensions. Learnable parameters introduced during network construction. It is set to 0 during network initialization. In the initial phase of model training, because... Not yet updated. The gating coefficient is always 1 to ensure that gradients can be fully backpropagated in the early stages of network training, prioritizing the establishment of brightness enhancement mappings to avoid the network falling into the "identity mapping trap" of outputting completely black images; as the model training process iterates and the backpropagation algorithm adaptively optimizes the parameters, The values ​​are gradually updated to positive values, so that the characteristic activation values ​​of dust-prone areas are smoothly decayed.

[0038] P2. Subsequently, the low-frequency component is gated. Input PG-SSM to perform long-distance dependency modeling and obtain the long-distance modeling output sequence.

[0039] Specifically, firstly, the two-dimensional spatial features Flattened into a one-dimensional feature sequence along a preset scanning path The input is mapped to the output using the discretized state transition formula:

[0040]

[0041] in, Represents implicit state. This represents the input one-dimensional feature sequence. Indicates the output. Here is the state transition matrix. and Let be the input projection matrix and the output projection matrix, respectively, and t represent the sequence index of the discretized one-dimensional feature.

[0042] To overcome the drawback of losing two-dimensional image spatial adjacency relationships in one-dimensional serialization, this invention employs a four-way cross-scanning strategy in PG-SSM. Specifically, the two-dimensional spatial features... Flatten the image along four different directions (top left to bottom right, bottom right to top left, top right to bottom left, and bottom left to top right) to generate four independent one-dimensional sequences. ,in These represent four scanning directions. After processing the four sequences using the discretized state transition formulas described above, long-distance modeling output sequences in the four directions are obtained. .

[0043]

[0044]

[0045] P3. Then, the output sequences in these four directions are reverse-flattened according to the original scanning direction to restore the original two-dimensional spatial resolution, and the channels are added to obtain the spatial fusion features. :

[0046] in, Indicates corresponding to the first The reverse two-dimensional feature recombination function is used in each scanning direction. Specifically, since the forward cross-scan flattens the two-dimensional spatial features into a one-dimensional sequence along four different directions, the reverse recombination function performs reciprocal geometric spatial mapping (including but not limited to the reverse arrangement of the one-dimensional sequence, the transpose and refolding of the tensor dimension) according to the corresponding scanning path, accurately restoring each feature element in the one-dimensional sequence to the original two-dimensional spatial coordinates, so as to ensure that the features in the four directions have strict physical spatial alignment when weighted and fused.

[0047] P4. Finally, it is connected to the unadjusted low-frequency component input to this module via a skip residual connection. Pixel-level summation yields the deep coding features of this layer. The formula is as follows:

[0048] The obtained deep coding features As the next network level (i.e., the first The input to the layer encoder continues to be passed down until the final layer of the encoder, resulting in the final deep-coded features; and the wavelet high-frequency components corresponding to each layer... These features are then retained as skip connection features and directly input into the corresponding decoder stage for subsequent high-frequency detail reconstruction.

[0049] S3D Feature Decoding. Each layer of decoding includes the following steps: S3D1. In the nth layer, the decoder first receives the low-frequency features of the input. The low-frequency features of this input are encoded as deep features in the final layer of the encoder in layer 1. The input features of each subsequent decoder are the output features of the previous decoder.

[0050] Simultaneously, extract the wavelet high-frequency components that are retained in parallel at the corresponding encoder level. (This corresponds to the sampling scale i), and these are concatenated and fused along the channel dimension to construct the original high-frequency features, denoted as . .

[0051] S3D2. Low-frequency features Compared with the original high-frequency features The input is fed into a priori-guided high-frequency enhancement module (PSRBlock) for feature enhancement. Since the high-frequency information in mine images is filled with dust particles, this invention also introduces an adaptive soft-gated filtering mechanism in this module. When using low-frequency information to correct high-frequency details, the corresponding level of noise prior feature map is explicitly utilized. As attention weights, the dust and noise responses in the high-frequency features are adaptively attenuated, and then cross-band structural information is incorporated to obtain the updated high-frequency features. The formula is as follows:

[0052] in, For cross-band feature transfer function (preferably in this embodiment) The specific function of convolution operation is to linearly map the channel dimension of low-frequency features to precisely align it with the channel dimension of high-frequency features, thereby achieving cross-band illumination and structural information supplementation. The element-wise multiplication of spatial and channel dimensions is used for a single-channel prior mask. During multiplication operations, copies are automatically broadcast along the channel dimension to precisely align the number of channels for high-frequency features; These are learnable parameters.

[0053] S3D3. Subsequently, the low-frequency characteristics With updated high-frequency features The data is stitched together along the channel dimension and then subjected to a two-dimensional inverse wavelet transform. Feature upsampling is performed to restore the spatial resolution to the previous scale, thus obtaining the decoded features of the current level. :

[0054] in, This indicates concatenation of channel dimensions.

[0055] Decoding features of this level This will serve as input to the next level of the decoder for subsequent decoding. After the entire decoder reconstructs the data layer by layer, the decoder outputs the final residual features (i.e., the final decoded features), which are then compared with the initial luminance components. Pixel-level addition is performed to obtain the enhanced luminance tensor. .

[0056] Preferably, to eliminate the problems caused by channel interleaving and merging in inverse wavelet reconstruction... The checkerboard artifact is addressed by upsampling the features and then immediately followed by a smooth convolutional layer with a stride of 1 and padding of 1. This smooth convolution layer, using a weighted average of pixel neighborhoods, forces the bridging of feature discontinuities, resulting in the smoothed decoded features of the current layer. :

[0057] In the formula, This indicates an anti-aliased smooth convolutional layer.

[0058] If this step is performed, then the smooth decoding feature of that level is used. It serves as the input to the next level of decoder for subsequent decoding.

[0059] S4. Finally, the enhanced brightness tensor With primary chromaticity components The images are re-stitched along the channel dimension, and the final enhanced image is output through an inverse color space transformation matrix. The formula is as follows:

[0060] Example 2: Additionally, as Figure 2 As shown, the present invention also provides an image enhancement system based on a priori guided wavelet state space model, including: a color space conversion module, a multi-scale prior pyramid generation module, a main enhancement branch, and a color space inverse conversion module.

[0061] The color space conversion module is used to convert native RGB low-light images. Convert to YCbCr color space and extract the luminance component of the Y channel. chromaticity components of the CbCr channel .

[0062] The multi-scale prior pyramid generation module extracts chromaticity components using a lightweight convolutional neural network (Noise-PEN). Basic prior mask probability map And the basic prior mask probability map Perform stepwise downsampling (DWT) to generate a multi-scale prior pyramid. .

[0063] The main enhancement branch is an encoder-decoder structure, with each encoder layer including a downsampling module (DWT) and a priori-guided low-frequency state-space model network module (PG-SSM). The downsampling module (DWT) performs a single feature downsampling on the input features using a two-dimensional discrete wavelet transform, decomposing the input into low-frequency components. and high frequency components { , , The prior-guided low-frequency state-space model network module (PG-SSM) processes the downsampled low-frequency components. Spatial fusion characteristics were obtained. and spatial integration features with low-frequency components Pixel-level summation yields the deep coding features of this layer. The deep coding features of this layer As input to the next network layer, the information continues to be passed down until the final layer of the encoder, resulting in the final deep encoded features.

[0064] Each decoder layer includes a priori guided high-frequency enhancement module (PSRBlock) and an upsampling module ( The priori-guided high-frequency enhancement module (PSRBlock) receives the input low-frequency characteristics. Simultaneously, high-frequency components at the corresponding scale are extracted. The features are then spliced ​​and fused along the channel dimension to construct the original high-frequency features. For the original high-frequency features Feature enhancement is performed to obtain updated high-frequency features. The formula is as follows: ,in, It is a cross-band characteristic transfer function. Represents element-wise multiplication of spatial and channel dimensions. For learnable parameters, For the noise prior feature maps of the corresponding scale in the multi-scale prior pyramid, the low-frequency features are... With updated high-frequency features The data is concatenated along the channel dimension, and then the upsampling module ( Two-dimensional inverse wavelet transform is used. Perform feature upsampling to obtain the decoded features of the current level. Decoding features at this level This data serves as input to the next level of the decoder for subsequent decoding. The decoder outputs the final decoded features, which are then compared with the initial luminance components. Pixel-level addition is performed to obtain the enhanced luminance tensor. , The color space inverse conversion module will enhance the luminance tensor. With primary chromaticity components The images are re-stitched along the channel dimension, and the final enhanced image is output through an inverse color space transformation matrix. .

[0065] Furthermore, this embodiment is merely a basic description of the image enhancement system based on the prior guided wavelet state space model of the present invention. Technical details not described in detail in this embodiment can be found in the methods provided in any embodiment of the present invention, and will not be repeated here.

[0066] Example 3: Those skilled in the art will clearly understand that the systems and methods of the above embodiments can be implemented using software plus necessary general-purpose hardware platforms. Of course, they can also be implemented using hardware, but in many cases, the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, node packaging device, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0067] Therefore, the present invention also provides an image enhancement device based on a priori guided wavelet state space model, comprising: a memory, a processor, and an image enhancement program based on a priori guided wavelet state space model stored in the memory and executable on the processor, wherein the image enhancement program based on a priori guided wavelet state space model is configured to implement an image enhancement method based on a priori guided wavelet state space model.

[0068] In addition, the present invention also provides a storage medium storing an image enhancement program based on a priori guided wavelet state-space model.

[0069] In reality, when deploying equipment or programs, a program may execute all steps or only one step, and all steps may be achieved through the cooperation of multiple programs. Therefore, the image enhancement program based on the prior guided wavelet state space model, when executed, implements all or a certain process of the image enhancement method based on the prior guided wavelet state space model.

[0070] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification, or direct or indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. An image enhancement method based on a priori guided wavelet state-space model, characterized in that, Includes the following steps: S1. Obtain the raw image It was then converted to the YCbCr color space, and the luminance component of the Y channel was extracted. and the chromaticity components of the CbCr channel ; S2. Using chromaticity components As input, a basic prior mask probability map is extracted using a convolutional neural network. And the basic prior mask probability map Perform stepwise downsampling to generate a multi-scale prior pyramid. Where i represents the scale; S3. Convert the brightness component The input is fed into the main enhancement branch for feature extraction and enhancement. The main enhancement branch adopts an encoder-decoder structure, and each layer of encoding includes the following steps: S3C1. Perform a single feature downsampling using a two-dimensional discrete wavelet transform on the input features, decomposing the input into low-frequency components. and high frequency components { , , The input features of the first layer are the luminance components. The input features of each subsequent layer are the output features of the previous layer. S3C2. Processing downsampled low-frequency components using a priori-guided low-frequency state-space model network PG-SSM. Spatial fusion characteristics were obtained. and spatial integration features with low-frequency components Pixel-level summation yields the deep coding features of this layer. The deep coding features of this layer The input is passed down to the next network layer until the final layer of the encoder, resulting in the final deep encoded features. Before scanning the state-space model network, the noise prior feature maps of the corresponding scale in the multi-scale prior pyramid are used as gating factors to perform element-wise attention weighting on the input features, thereby suppressing the feature intensity of high-noise regions. Step S3C2 includes: P1. Extracting noisy prior feature maps of the corresponding scale from the multi-scale prior pyramid. For low-frequency components Adaptive soft-gating is performed to obtain the gated low-frequency component. The formula is as follows: in, For linear mapping operations, For element-wise multiplication of spatial and channel dimensions, For learnable parameters, P2. Gating low-frequency components Along different directions of the image Each part is flattened to generate a one-dimensional sequence. And through implicit states Input Mapped to output The formula is as follows: in, Represents implicit state. This represents the input one-dimensional feature sequence. Indicates the output. Here is the state transition matrix. and Let be the input projection matrix and the output projection matrix, respectively, and t represent the sequence index of the discretized one-dimensional feature. P3. Perform an inverse flattening operation on the output sequence in each direction according to the original scanning direction to restore the original two-dimensional spatial resolution, and then sum the channels to obtain the spatial fusion features. , P4. Integrating spatial features with low-frequency components Pixel-level summation yields the deep coding features of this layer. The deep coding features obtained As input to the next network layer, it continues to be passed down until the final layer of the encoder, resulting in the final deep encoded features; Each layer of decoding includes the following steps: S3D1. Decoder receives low-frequency characteristics of the input. The low-frequency features of this input are encoded as deep features in the final layer of the encoder in the first layer. The input features of each subsequent decoder layer are the output features of the previous decoder layer. At the same time, high-frequency components at the corresponding scale are extracted. The features are then spliced ​​and fused along the channel dimension to construct the original high-frequency features. , S3D2. Original high-frequency features Feature enhancement is performed to obtain updated high-frequency features. The formula is as follows: in, For cross-band characteristic transfer function, Represents element-wise multiplication of spatial and channel dimensions. For learnable parameters, This represents the noise prior feature map at the corresponding scale in the multi-scale prior pyramid. S3D3. Low-frequency features With updated high-frequency features The data is stitched together along the channel dimension and then subjected to a two-dimensional inverse wavelet transform. Perform feature upsampling to obtain the decoded features of the current level. Decoding features at this level This data serves as input to the next level of the decoder for subsequent decoding. The decoder outputs the final decoded features, which are then compared with the initial luminance components. Pixel-level addition is performed to obtain the enhanced luminance tensor. , S4. Enhance the brightness tensor With primary chromaticity components The images are stitched together along the channel dimension, and the final enhanced image is obtained through inverse color space transformation. .

2. The image enhancement method based on a priori guided wavelet state-space model as described in claim 1, characterized in that, Step S3D3 obtains the decoding features of the current level. Then, calculate the smooth decoding features of the current level. The formula is as follows: In the formula This represents the anti-aliased smooth convolution computation and uses the smooth decoding features of this level. It serves as the input to the next level of decoder for subsequent decoding.

3. The image enhancement method based on a priori guided wavelet state-space model as described in claim 1, characterized in that, It is set to 0 during network initialization.

4. An image enhancement system based on a priori guided wavelet state-space model, implementing the method described in any one of claims 1-3, characterized in that, include: The system includes a color space conversion module, a multi-scale prior pyramid generation module, a main enhancement branch, and a color space inverse conversion module. The main enhancement branch is an encoder-decoder structure. Each encoder layer includes a downsampling module (DWT) and a priori-guided low-frequency state-space model network module (PG-SSM). Each decoder layer includes a priori-guided high-frequency enhancement module (PSR BLOCK) and an upsampling module (IWT).

5. An image enhancement device based on a priori guided wavelet state-space model, characterized in that, include: The image enhancement program based on a priori guided wavelet state space model, stored in the memory and executable on the processor, is configured to implement the image enhancement method based on a priori guided wavelet state space model as described in any one of claims 1-3.

6. A storage medium, characterized in that, The storage medium stores an image enhancement program based on a priori guided wavelet state space model, which, when executed, implements the image enhancement method based on a priori guided wavelet state space model as described in any one of claims 1-3.

Images