Image defogging enhancement method for mine coal face in dust fog environment

Abstract

The invention relates to an image defogging enhancement method for a mine coal face in a dust fog environment. The method comprises the following steps: carrying out priori extraction of four kinds of statistics on an input mine dust fog image, a mine dust fog depth map and a mine real dust fog-free image; depth information prior remapping and real fog-free information prior remapping are carried out, four kinds of statistic distribution information of a depth image and a real fog-free image are utilized, feature distribution of a haze image is dynamically adjusted, physical prior understanding of a network on the depth image is improved, and residual fog color difference is relieved; in a u-net network framework, plug-and-play directional wavelet convolution is utilized, directional selection enhancement processing is carried out on four sub-bands, the direction sensing capacity of wavelet convolution is improved, and the performance is improved. The method is suitable for the mine dust fog environment, the clear high-quality image effect is achieved while defogging is achieved, reliable visual input is provided for subsequent intelligent analysis tasks, the intelligent monitoring level of a coal mine is improved, and the recognition capacity of potential safety hazards of the coal mine is enhanced.

Description

A method for image dehazing and enhancement in dusty and foggy environments at coal mining faces Technical Field

[0001] This invention relates to a method for image dehazing and enhancement in dusty and foggy environments at coal mining faces, belonging to the field of image dehazing technology. Background Technology

[0002] Image dehazing, a core underlying task in the field of computer vision, aims to improve image quality degradation caused by factors such as haze, dust, and water mist, thereby enhancing the visual perception of images. In the underground working environment of coal mines, coal ash and dust generated by long-term production operations are continuously dispersed in the air. In addition, the widespread deployment of dust suppression spraying systems in mine working areas results in the underground air being filled with high concentrations of dust and fog. At the same time, the lighting conditions inside the mine are complex, with various lighting forms such as direct light from mine lamps, diffuse reflection from LED lighting, and secondary reflection from reflective walls intertwined and superimposed. These factors together lead to a serious deterioration in the image quality captured by the imaging equipment, specifically manifested in the following three aspects: (1) Low contrast: Coal dust particles have a significant absorption effect on light, resulting in an overall dark image and blurred edges of the target object; (2) Residual fog in local areas: The droplets generated in the dust suppression spraying area are relatively large, which greatly enhances the scattering effect of light, causing a significant decrease in visibility in these local areas; (3) High dynamic range and non-uniform fog distribution: The visibility varies greatly in different areas underground. Some areas are almost free of fog interference, while other areas are completely covered by dense fog. This characteristic makes it difficult for traditional image defogging methods to achieve full-scene adaptation.

[0003] The above problems directly affect the monitoring camera system and intelligent detection system in the mine, seriously interfering with the accuracy of tasks such as target detection, personnel identification, and equipment monitoring, and failing to meet the safety requirements of coal mining. Summary of the Invention

[0004] The purpose of this invention is to provide an image defogging and enhancement method for coal mining faces under dust and fog conditions. This method can adapt to the dust and fog environment of mines, achieve defogging while obtaining clear and high-quality image effects, provide reliable visual input for subsequent intelligent analysis tasks, improve the level of intelligent monitoring in coal mines, and enhance the ability to identify coal mine safety hazards.

[0005] To achieve the above objectives, this invention provides an image dehazing enhancement method for coal mining faces under dust and fog conditions, comprising the following steps: S1, Prior Extraction and Preliminary Processing: Four statistical parameters are extracted prior from the input coal mining dust and fog image, coal mining dust and fog depth map, and a real coal mining image without dust and fog; S2, Prior-Driven Feature Remapping: This includes prior remapping of depth information and prior remapping of real fog-free information. Utilizing the distribution information of the four statistical parameters from the depth map and the real fog-free image, the feature distribution of the haze image is dynamically adjusted to improve the network's physical prior understanding of the depth map, thereby alleviating residual fog. Color difference; the approach primarily uses mean-variance remapping, with skewness and kurtosis compensation for feature fine-tuning; remapping is performed in two stages. The first stage dynamically adjusts the statistical feature distribution of the mine dust and fog image by analyzing the statistical distribution of the real dust-free image, making the image restoration effect closer to the real image and reducing artifacts and detail loss. The second stage dynamically adjusts the statistical feature distribution of the mine dust and fog image by analyzing the statistical distribution of the mine dust and fog depth map. This stage aims to enable the network to truly understand the prior information on dust and fog concentration distribution provided by the depth map, thereby achieving efficient multimodal dehazing. Finally, the feature map remapping weights provided by the depth map and the fog-free image are dynamically weighted to achieve feature remapping of the haze image, i.e., adjusting the statistical distribution; S3, in the u-net network framework, plug-and-play directional wavelet convolution is used. By performing directional selection enhancement processing on the four sub-bands, the wavelet convolution's ability to perceive direction is improved, thereby improving performance.

[0006] Furthermore, the mine dust and fog image, mine dust and fog depth map, and actual dust-free image in S1 correspond to the haze feature map, haze image depth feature map, and actual haze-free feature map, respectively, and the four statistical measures are the mean, mean, and so on. Standard deviation skewness and kurtosis The calculation formulas are as follows: ; ; ; Where h and w represent row index and column index respectively, indicating a row and column in an image or matrix; This represents the element located in row h and column w; H and W represent the height and width of the image, respectively. It is the total element of the image.

[0007] Furthermore, the specific process of S2 is as follows: S2.1, for the prior remapping of depth information, let the four statistical measures corresponding to the haze image and the haze depth map be respectively... , ; , ; , ; , Haze image characteristics are denoted as First, the input features, i.e., the haze images, are normalized for mean and variance. The normalized features are denoted as follows: : Next, the normalized features are remapped to the mean of the target features. The haze features remapped based on the mean and variance of the haze depth map are denoted as follows: : S2.2, skewness and kurtosis processing: Compensation coefficients are calculated using a fully connected network. The channel mean values ​​of the four statistics for each sample depth map are calculated, and the channel mean values ​​are concatenated and recorded as follows: Then, through the fully connected layer... The compensation coefficients are converted into skewness and kurtosis, used to adjust the skewness and kurtosis of the input features; let the compensation coefficients for skewness and kurtosis after the fully connected layer be denoted as . and The compensation terms for skewness and kurtosis are denoted as , : ; The compensation term is remapped onto the normalized features of the haze image, and the feature map adjusted for skewness and kurtosis is denoted as: S2.3, The haze image features obtained after deep prior-guided remapping are denoted as... : S2.4. For the prior remapping of true haze-free information, let the mean, variance, skewness, and kurtosis of the true haze-free map be respectively... , , and Let the haze characteristics after remapping guided by the mean and variance of the real haze-free map be denoted as . : S2.5, skewness, and kurtosis are handled by calculating compensation coefficients using a fully connected network. The channel means of the four statistics for each sample's true haze-free image are calculated, and the channel means are concatenated and denoted as... Then, through the fully connected layer... The compensation coefficients are converted into skewness and kurtosis, used to adjust the skewness and kurtosis of the input features; let the compensation coefficients for skewness and kurtosis after the fully connected layer be denoted as . and The compensation terms for skewness and kurtosis are denoted as , : ; The compensation term obtained from the real haze-free image is remapped onto the normalized features of the haze image, and the feature map adjusted for skewness and kurtosis is denoted as... : S2.6, The haze features after remapping guided by the no-haze prior are denoted as... : S2.7. After balancing the weight contributions of the two branches—the fog-free prior-guided remapping and the deep prior-guided remapping—by the gating weighting module, the final output is denoted as... The gating weights after remapping guided by fog-free priors and remapping guided by deep priors are denoted as follows: and The specific steps are as follows: ; ; ; ;in, and This represents a size of and Two-dimensional convolution; This represents average pooling; It is a non-linear activation function. Represents channel division, then Represented as: Furthermore, in S3, the four sub-bands are a vertical sub-band, a horizontal sub-band, a low-frequency sub-band, and a diagonal sub-band. The specific process of performing directional selection enhancement processing on each sub-band is as follows: S3.1, For the diagonal sub-band, in order to further enhance the directional awareness capability of the diagonal sub-band, a rotational convolution denoted as Roconv2d is designed to create a coordinate network for each pixel of the convolution. Rotation is achieved by changing the coordinates of each pixel. Let the coordinates of each pixel be... The rotated coordinates are The rotation angle is denoted as : ; The true rotation of the convolution is achieved by mapping the rotated coordinate points to new sampling points to generate a new convolution; given that the diagonal subbands of wavelet convolution are in and It has strong feature expression, and is designed separately. and The rotational convolution operates on the diagonal subband; to balance the weighted contribution of each angle. Because it is extremely sensitive to directionality, traditional gated convolution destroys directional texture information. An angle-controlled mechanism is introduced, the core of which is to teach it the mapping relationship between angles and weights, allowing for more refined and sensitive information. The specific operation is as follows: First, the input vector is converted into a normalized vector: in, Next, calculate the similarity between the synthesized vector and the angle: ; Calculate their angular similarity: Angular similarity is calculated by calculating the input vector and the composite vector Cosine similarity: ;angle Calculated using the inverse cosine function: Ultimately, a hybrid weighting system using Gaussian and MLP methods is employed to learn the mapping relationship between angles and weights. Let the Gaussian weights be... MLP weights are The mixed weights are : ; For each input vector Increase the weights to enhance them: According to the weighting coefficients Perform weighted summation, using Weight normalization: Finally, we obtain the weighted output of the diagonal subband, and denote the direction-enhanced diagonal subband as... : S3.2, Enhance the vertical subband: Let the vertical subband of wavelet decomposition be . The subband with enhanced vertical subband directionality is denoted as : ; ; ; ;in, Represents intermediate transitional variables in the formula. It is a one-dimensional vertical stripe depthwise convolution with a size of , It is the size of Two-dimensional depthwise convolution, S3.3, Enhancement of the horizontal subband: Let the horizontal subband of wavelet decomposition be (where S is a nonlinear activation function). The subband after the horizontal subband's directionality is enhanced is denoted as : ; ; ; ; ;in, The intermediate transition variable represents the above formula; S3.4, for the low-frequency sub-band of wavelet decomposition, multi-scale feature extraction is first performed, and the four statistical measures are enhanced. Then, the SE channel is used to further enhance the channel representation. Specifically, let the low-frequency sub-band of wavelet decomposition be... The enhanced low-frequency subband is : ;in, It is a non-linear activation function. It is a depthwise separable convolution with a size of ;right Four statistical measures were calculated and the intensity was dynamically adjusted. The four statistical measures of Z are denoted as follows: , , and , Representative to Four statistical measures are calculated, and their output is denoted as follows: : ; S3.5. Reassemble the features of the four sub-bands, i.e., concatenate them into a wavelet transform tensor: ;in, This represents the concatenated wavelet transform tensor. This indicates that concatenation is performed along the channel dimension; the final directional wavelet convolution output is denoted as: Where IWT is the inverse wavelet transform.

[0008] This invention extracts four statistical priors from input images of mine dust and fog, depth maps of mine dust and fog, and real dust-free and fog-free images of mines. Then, it employs a dual-priority-driven remapping method using depth priors and real fog-free priors to dynamically adjust the feature distribution of the fog image, improving the network's physical prior understanding of the depth map and thus mitigating residual fog color differences. Finally, within the u-net network framework, it utilizes plug-and-play directional wavelet convolution, enhancing the wavelet convolution's directional perception capability through directional selection enhancement processing on four sub-bands. This invention achieves clear, high-quality image results while removing fog, providing reliable visual input for subsequent intelligent analysis tasks, improving the level of intelligent monitoring in coal mines, and enhancing the ability to identify coal mine safety hazards. Attached Figure Description

[0009] Figure 1 is a flowchart of the method of the present invention; Figure 2 is a schematic diagram of the prior-driven remapping structure of the present invention; Figure 3 is a schematic diagram of the Dynamic Weight Generator structure of the present invention; Figure 4 is an experimental diagram of the directional wavelet convolution motivation of the present invention; Figure 5 is a schematic diagram of the vertical sub-band enhancement module structure of the present invention; Figure 6 is a schematic diagram of the horizontal sub-band enhancement module structure of the present invention; Figure 7 is a schematic diagram of the low-frequency enhancement module structure of the present invention; Figure 8 is a schematic diagram of the SE channel module structure of the present invention. Detailed Implementation

[0010] The invention will now be further described with reference to the accompanying drawings.

[0011] As shown in Figure 1, an image dehazing enhancement method for coal mining faces under dust and fog conditions includes the following steps: S1, Prior Extraction and Preliminary Processing: Four statistical parameters are extracted prior from the input coal mining dust and fog image, coal mining dust and fog depth map, and a real coal mining image without dust and fog; S2, Prior-Driven Feature Remapping: This includes prior remapping of depth information and prior remapping of real fog-free information. Using the distribution information of the four statistical parameters in the depth map and the real fog-free image, the feature distribution of the haze image is dynamically adjusted to improve the network's physical prior understanding of the depth map, thereby alleviating residual fog color differences; The formula primarily uses mean-variance remapping, with skewness and kurtosis compensation for feature fine-tuning. Remapping is performed in two stages. The first stage dynamically adjusts the statistical feature distribution of the mine dust and fog image by analyzing the statistical distribution of the real dust-free image, making the image restoration effect closer to the real image and reducing artifacts and detail loss. The second stage dynamically adjusts the statistical feature distribution of the mine dust and fog image by analyzing the statistical distribution of the mine dust and fog depth map. This stage aims to enable the network to truly understand the prior information on dust and fog concentration distribution provided by the depth map, thereby achieving efficient multimodal dehazing. Finally, the feature map remapping weights provided by the depth map and the fog-free image are dynamically weighted to achieve feature remapping of the haze image, i.e., adjusting the statistical distribution. S3. In the u-net network framework, plug-and-play directional wavelet convolution is used. Directional selection enhancement processing on the four sub-bands improves the wavelet convolution's ability to perceive direction, thereby improving performance.

[0012] Mine dust and fog images, mine dust and fog depth maps, and actual dust-free and fog-free mine images correspond to haze feature maps, haze image depth feature maps, and actual fog-free feature maps, respectively. The four statistical measures are the mean, mean, and mean values. Standard deviation skewness and kurtosis The calculation formulas are as follows: ; ; ; Where h and w represent row index and column index respectively, indicating a row and column in an image or matrix; This represents the element located in row h and column w; H and W represent the height and width of the image, respectively. It is the total element of the image.

[0013] As shown in Figure 2, in a preferred implementation, the prior-driven feature remapping process is as follows: S2.1, for the prior remapping of depth information, let the four statistical measures corresponding to the haze image and the haze depth map be respectively... , ; , ; , ; , Haze image characteristics are denoted as First, the input features, i.e., the haze images, are normalized for mean and variance. The normalized features are denoted as follows: : Next, the normalized features are remapped to the mean of the target features. The haze features remapped based on the mean and variance of the haze depth map are denoted as follows: : S2.2, skewness and kurtosis processing: Compensation coefficients are calculated using a fully connected network. The channel mean values ​​of the four statistics for each sample depth map are calculated, and the channel mean values ​​are concatenated and recorded as follows: Then, through the fully connected layer... The compensation coefficients are converted into skewness and kurtosis, used to adjust the skewness and kurtosis of the input features; let the compensation coefficients for skewness and kurtosis after the fully connected layer be denoted as . and The compensation terms for skewness and kurtosis are denoted as , : ; The compensation term is remapped onto the normalized features of the haze image, and the feature map adjusted for skewness and kurtosis is denoted as... : S2.3, The haze image features obtained after deep prior-guided remapping are denoted as... : S2.4. For the prior remapping of true haze-free information, let the mean, variance, skewness, and kurtosis of the true haze-free map be respectively... , ,and Let the haze characteristics after remapping guided by the mean and variance of the real haze-free map be denoted as . : S2.5, skewness, and kurtosis are handled by calculating compensation coefficients using a fully connected network. The channel means of the four statistics for each sample's true haze-free image are calculated, and the channel means are concatenated and denoted as... Then, through the fully connected layer... The compensation coefficients are converted into skewness and kurtosis, used to adjust the skewness and kurtosis of the input features; let the compensation coefficients for skewness and kurtosis after the fully connected layer be denoted as . and The compensation terms for skewness and kurtosis are denoted as , : ; The compensation term obtained from the real haze-free image is remapped onto the normalized features of the haze image, and the feature map adjusted for skewness and kurtosis is denoted as... : S2.6, The haze features after remapping guided by the no-haze prior are denoted as... : S2.7, As shown in Figure 3, after the gated weighting module balances the weight contributions of the two branches—the fog-free prior-guided remapping and the depth prior-guided remapping—the final output is denoted as... The gating weights after remapping guided by fog-free priors and remapping guided by deep priors are denoted as follows: and The specific steps are as follows: ; ; ; ;in, and This represents a size of Two-dimensional convolution of sum; This represents average pooling; It is a non-linear activation function. Represents channel division, then Represented as: .

[0014] As shown in Figure 4, the four sub-bands are a vertical sub-band, a horizontal sub-band, a low-frequency sub-band, and a diagonal sub-band. The specific process of directional selection enhancement for each sub-band is as follows: S3.1. For the diagonal sub-band, in order to further enhance the directional awareness of the diagonal sub-band, a rotational convolution denoted as Roconv2d is designed to create a coordinate network for each pixel of the convolution. Rotation is achieved by changing the coordinates of each pixel. Let the coordinates of each pixel be... The rotated coordinates are The rotation angle is denoted as : ; The true rotation of the convolution is achieved by mapping the rotated coordinate points to new sampling points to generate a new convolution; given that the diagonal subbands of wavelet convolution are in and It has strong feature expression, and is designed separately. and The rotational convolution operates on the diagonal subband; to balance the contribution of each weighted angle, because it is extremely sensitive to directionality, traditional gated convolution would destroy directional texture information. Therefore, an angle-controlled mechanism is introduced. The core is to teach it the mapping relationship between angles and weights, allowing for more refined and sensitive information. The specific operation is as follows: First, the input vector is converted into a normalized vector: in, Next, calculate the similarity between the synthesized vector and the angle: ; Calculate their angular similarity: Angular similarity is calculated by calculating the input vector and the composite vector Cosine similarity: ;angle Calculated using the inverse cosine function: Ultimately, a hybrid weighting system using Gaussian and MLP methods is employed to learn the mapping relationship between angles and weights. Let the Gaussian weights be... MLP weights are The mixed weights are : ; For each input vector Increase the weights to enhance them: According to the weighting coefficients Perform weighted summation, using Weight normalization: Finally, we obtain the weighted output of the diagonal subband, and denote the direction-enhanced diagonal subband as... : S3.2, As shown in Figure 5, the vertical subband is enhanced: Let the vertical subband of the wavelet decomposition be . The subband with enhanced vertical subband directionality is denoted as : ; ; ; ; ;in, The intermediate variables in the above formula (without special meaning) represent the intermediate variables. It is a one-dimensional vertical stripe depthwise convolution with a size of , It is the size of Two-dimensional depthwise convolution, The activation function is nonlinear; S3.3, as shown in Figure 6, enhances the horizontal sub-band: Let the horizontal sub-band of wavelet decomposition be... The subband after the horizontal subband's directionality is enhanced is denoted as : ; ; ; ; ;in, The intermediate transition variables represented in the above formula (without special meaning); S3.4, as shown in Figure 7, for the low-frequency sub-band of wavelet decomposition, multi-scale feature extraction is first performed, and the four statistical quantities are enhanced, as shown in Figure 8. Then, the SE channel is used to further enhance the channel representation, specifically: Let the low-frequency sub-band of wavelet decomposition be... The enhanced low-frequency subband is : ;in, It is a non-linear activation function. It is a depthwise separable convolution with a size of ;right Four statistical measures were calculated and the intensity was dynamically adjusted. The four statistical measures of Z are denoted as follows: , , and , Representative to Four statistical measures are calculated, and their output is denoted as follows: : ; S3.5. Reassemble the features of the four sub-bands, i.e., concatenate them into a wavelet transform tensor: ;in, This represents the concatenated wavelet transform tensor. This indicates concatenation along the channel dimension; the final directional wavelet convolution output is denoted as... : Where IWT is the inverse wavelet transform.

[0015] At The performance of different models was compared on the dataset, and the results are shown in Table 1: Table 1 Comparison table of performance results of different models on different datasets The data above shows that the method provided by this invention has better defogging enhancement performance than other existing image defogging methods and is more suitable for dusty and foggy environments in mines.

Claims

1. A method for image dehazing and enhancement in a dusty and foggy environment at a coal mining face, characterized in that, The process includes the following steps: S1. Prior extraction and preliminary processing: Four statistical parameters are extracted from the input mine dust and fog image, mine dust and fog depth map, and real dust-free and fog-free mine image; S2. Prior-driven feature remapping: This includes prior remapping of depth information and prior remapping of real fog-free information. By utilizing the distribution information of the four statistical parameters of the depth map and the real fog-free image, the feature distribution of the haze image is dynamically adjusted to improve the network's physical prior understanding of the depth map, thereby alleviating the color difference caused by residual fog. The method is mainly based on mean-variance remapping, with skewness and kurtosis compensation for feature fine-tuning; S3. In the u-net network framework, plug-and-play directional wavelet convolution is used to enhance the wavelet convolution's ability to perceive direction by performing directional selection enhancement processing on the four sub-bands.

2. The image dehazing and enhancement method for coal mining faces under dust and fog conditions according to claim 1, characterized in that, The mine dust and fog image, mine dust and fog depth map, and mine real dust-free fog image in S1 correspond to the haze feature map, haze image depth feature map, and real fog-free feature map, respectively. The four statistical measures are the mean, mean, and so on. Standard deviation skewness and kurtosis The calculation formulas are as follows: ； ； ； Where h and w represent row index and column index respectively, indicating a row and column in an image or matrix; This represents the element located in row h and column w; H and W represent the height and width of the image, respectively. It is the total element of the image.

3. The image dehazing and enhancement method for coal mining faces under dust and fog conditions according to claim 2, characterized in that, The specific process of S2 is as follows: S2.1, For the prior remapping of depth information, let the four statistical measures corresponding to the haze image and the haze depth map be respectively... ， ； ， ； ， ； ， The characteristics of haze images are denoted as: First, the input features, i.e., the haze images, are normalized for mean and variance. The normalized features are denoted as follows: ： Next, the normalized features are remapped to the mean of the target features. The haze features remapped based on the mean and variance of the haze depth map are denoted as follows: : ； S2.2, skewness and kurtosis processing: Compensation coefficients are calculated using a fully connected network. The channel means of the four statistics for each sample depth map are calculated, and the channel means are concatenated and denoted as... ； Then, after passing through the fully connected layer... The compensation coefficients are converted into skewness and kurtosis, used to adjust the skewness and kurtosis of the input features; let the compensation coefficients for skewness and kurtosis after the fully connected layer be denoted as . and The compensation terms for skewness and kurtosis are denoted as 、 ： ； The compensation term is remapped onto the normalized features of the haze image, and the feature map adjusted for skewness and kurtosis is denoted as... : ； S2.3, The haze image features obtained after deep prior-guided remapping are denoted as... ： S2.

4. For the prior remapping of true haze-free information, let the mean, variance, skewness, and kurtosis of the true haze-free map be respectively... 、 、 and Let the haze characteristics after remapping guided by the mean and variance of the real haze-free map be denoted as . ： ； S2.5, skewness and kurtosis processing: Compensation coefficients are calculated using a fully connected network. The channel means of the four statistics for each sample's true haze-free image are calculated, and the channel means are concatenated and denoted as... ； Then, after passing through the fully connected layer... The compensation coefficients are converted into skewness and kurtosis, used to adjust the skewness and kurtosis of the input features; let the compensation coefficients for skewness and kurtosis after the fully connected layer be denoted as . and The compensation terms for skewness and kurtosis are denoted as 、 ： ； The compensation term obtained from the real haze-free image is remapped onto the normalized features of the haze image, and the feature map adjusted for skewness and kurtosis is denoted as... ： S2.6, The haze features after remapping guided by the no-haze prior are denoted as... ： ； S2.

7. After balancing the weight contributions of the two branches—the fog-free prior-guided remapping and the deep prior-guided remapping—using the gating weighting module, the final output is denoted as... The gating weights after remapping guided by fog-free priors and remapping guided by deep priors are denoted as follows: and The specific steps are as follows: ； ； ； ;in, and This represents a size of and Two-dimensional convolution; This represents average pooling; It is a non-linear activation function. Represents channel division, then Represented as: 。 4. The image dehazing and enhancement method for coal mining faces under dust and fog conditions according to claim 3, characterized in that, The four sub-bands in S3 are a vertical sub-band, a horizontal sub-band, a low-frequency sub-band, and a diagonal sub-band. The specific process of directional selection enhancement processing for each sub-band is as follows: S3.1, For the diagonal sub-band, a rotational convolution denoted as Roconv2d is designed to create a coordinate network for each pixel of the convolution. Rotation is achieved by changing the coordinates of each pixel. Let the coordinates of each pixel be... The rotated coordinates are The rotation angle is denoted as ： ； True convolution rotation is achieved by mapping the rotated coordinates to new sampling points to generate a new convolution; respectively... and The rotational convolution operates on the diagonal subband; to balance the contribution of each weighted angle, an angle control mechanism is introduced, specifically as follows: First, the input vector is transformed into a normalized vector: in, Next, calculate the similarity between the synthesized vector and the angle: ; Calculate their angular similarity: Angular similarity is calculated by input vector and the composite vector Cosine similarity: ;angle Calculated using the inverse cosine function: Finally, a hybrid weighting of Gaussian and MLP is applied to the channels, with the Gaussian weight set to [value missing]. MLP weights are The mixed weights are ： ； For each input vector Increase the weights to enhance them: According to the weighting coefficients Perform weighted summation, using Weight normalization: Finally, we obtain the weighted output of the diagonal subband, and denote the direction-enhanced diagonal subband as... ： S3.2, Enhance the vertical subband: Let the vertical subband of wavelet decomposition be . The subband with enhanced vertical subband directionality is denoted as ： ； ； ； ； ;in, Represents intermediate transitional variables in the formula. It is a one-dimensional vertical strip depthwise convolution with a size of ， It is the size of Two-dimensional depthwise convolution, S3.3, Enhancement of the horizontal subband: Let the horizontal subband of wavelet decomposition be (where S is a nonlinear activation function). The subband after the horizontal subband's directionality is enhanced is denoted as ： ； ； ； ； ;in, The intermediate transition variable represents the above formula; S3.4, for the low-frequency sub-band of wavelet decomposition, first perform multi-scale feature extraction, enhance the four statistical quantities, and then further enhance the channel representation using the SE channel, specifically: let the low-frequency sub-band of wavelet decomposition be... The enhanced low-frequency subband is ： ;in, It is a non-linear activation function. It is a depthwise separable convolution with a size of ;right Four statistical measures were calculated and the intensity was dynamically adjusted. The four statistics are denoted as follows: 、 、 and ， Representative to Four statistical measures are calculated, and their output is denoted as follows: ： ； S3.

5. Reassemble the features of the four sub-bands, i.e., concatenate them into a wavelet transform tensor: ;in, This represents the concatenated wavelet transform tensor. This indicates concatenation along the channel dimension; the final directional wavelet convolution output is denoted as... ： Where IWT is the inverse wavelet transform.

Images