Physical prior knowledge-based severe weather image enhancement method and system

Abstract

The invention discloses a physical priori knowledge-based severe weather image enhancement method and system, and relates to the technical field of image enhancement, and the method comprises the steps: inputting a to-be-enhanced image into a trained image enhancement model, and outputting a semantic segmentation image; the training process of the image enhancement model comprises the following steps: acquiring images in different driving environments in sunny days and severe weather, and constructing a training data set; using the sunny day image as a reference image, learning a distribution difference between the sunny day image and the severe weather image based on a sharpness module, separating a weather component from the severe weather image, and generating a sharpness image; physical prior is introduced into the image enhancement module, image enhancement processing is carried out on the sharp image, and an enhanced image is obtained; performing semantic segmentation on the enhanced image to obtain a semantic segmentation image with detail features; according to the severe weather image enhancement method and system, the definition and quality of rain and snow images are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image enhancement, and in particular to a method and system for enhancing images in severe weather based on physical prior knowledge. Background Art

[0002] In the field of autonomous driving, the reliability of visual sensors under different climatic conditions is crucial. However, rainy and snowy weather can significantly reduce the quality of images. The images collected by visual sensors are often disturbed by factors such as raindrops, snowflakes, and haze, which causes the clarity, brightness, contrast, and hue of the images to shift. These weather factors not only affect the overall visual effect of the image, but also affect the accuracy of subsequent computer vision tasks, especially semantic segmentation, object detection, and obstacle recognition, which are directly related to safety.

[0003] Traditional image enhancement techniques usually process rainy and snowy images through simple filtering, denoising or brightness adjustment, but these methods often cannot fundamentally remove the characteristic components introduced by rainy and snowy weather, and it is difficult to achieve adaptive enhancement under different weather conditions. Current research has gradually begun to explore the use of deep learning technology combined with physical prior knowledge to achieve image enhancement in rainy and snowy weather. In particular, by designing a dedicated loss function, the model can perform self-supervised learning by referring to sunny images without a paired data set to generate clearer and interference-free images. However, these methods still have room for improvement in efficiency and accuracy. Summary of the invention

[0004] Based on the technical problems existing in the background technology, the present invention proposes a severe weather image enhancement method and system based on physical prior knowledge, which significantly improves the clarity and quality of rain and snow images, enhances the visual perception capability of the automatic driving system under rain and snow conditions, and ensures driving safety.

[0005] The present invention proposes a severe weather image enhancement method and system based on physical prior knowledge, which inputs the image to be enhanced into a trained image enhancement model and outputs a semantic segmentation image;

[0006] The training process of the image enhancement model is as follows:

[0007] Obtain images of different driving environments in sunny and bad weather to build a training data set;

[0008] Using the sunny image as the reference image, the clearing module learns the distribution difference between the sunny image and the bad weather image, separates the weather component from the bad weather image, and generates a clear image.

[0009] Introducing physical priors into the image enhancement module, the image enhancement processing is performed on the cleared image to obtain the enhanced image;

[0010] Perform semantic segmentation on the enhanced image to obtain a semantic segmentation image with detailed features;

[0011] Construct a total loss function and adjust the parameters of the image enhancement model.

[0012] Further, the sharpening module includes six sequentially connected convolutional layers;

[0013] The first three convolutional layers are used to process the severe weather image and downsample layer by layer to reduce the size of the severe weather image and extract low-level features of the severe weather image to obtain the first image;

[0014] The last three convolutional layers gradually upsample the first image through deconvolution operations, restore the first image to its original resolution, and ensure that the low-level features in the first image are preserved, thereby generating a sharpened image;

[0015] Batch regularization is applied after each convolutional layer to stabilize the feature distribution, and the ReLU activation function is used to enhance the nonlinear expression ability of the clarity module.

[0016] Further, the image enhancement module includes six convolutional layers connected in sequence;

[0017] The convolution operation of each convolution layer gradually doubles the number of channels and reduces the spatial size to half of the previous convolution layer to extract the multi-scale feature information of the clear image step by step. The output of the last convolution layer of the image enhancement module generates the parameters required for image enhancement; the extracted clear image is enhanced based on the required parameters.

[0018] Furthermore, physical priors are introduced into the image enhancement module to perform image enhancement processing on the cleared image, specifically:

[0019] A model was established based on atmospheric scattering theory to describe the scattering process of light in the atmosphere under severe weather conditions, and the effects of haze and suspended water molecules on the clear image were obtained;

[0020] Gamma correction technology is used to adjust the brightness and contrast of the cleared image to compensate for the brightness reduction and contrast shift caused by bad weather, thereby obtaining an enhanced image.

[0021] Furthermore, the atmospheric scattering theory model is specifically as follows:

[0022] I(x)=J(x)t(x)+A(1-t(x));

[0023] Among them, I(x) is the pixel value of the severe weather image at position x, J(x) is the pixel value of the real image of the scene at position x, A represents the background brightness, and t(x) is the transmittance, which is defined as the light transmission rate from the scene to the camera.

[0024] Furthermore, gamma correction technology is used to adjust the brightness and contrast of the clear image, specifically:

[0025]

[0026] Among them, I in is the pixel value of the input severe weather image, I out is the pixel value of the enhanced image, γ is the gamma value, when γ<1, it is used to enhance the dark details, and when γ>1, it is used to compress the bright details.

[0027] Furthermore, the total loss function includes clarity loss, exposure loss and semantic segmentation loss.

[0028] Further, the clarity loss includes a priori loss and a structural similarity index loss;

[0029] The prior loss L ref for:

[0030]

[0031] The structural similarity index loss L ssim for:

[0032]

[0033] The exposure loss L exp for:

[0034]

[0035] The semantic segmentation loss L seg for:

[0036]

[0037] Among them, C(I ref ) is the cleared image output by the clearing module, is the all-0 feature map, ∥·∥2 is the binary norm, SSIM is the structural index, C in is the cleared image output by the clearing module, I in For severe weather images, I ref is the reference image, s is a set constant to measure the similarity between the reference image and the severe weather image, ||·||1 is a norm, M is the number of local regions divided by the enhanced image, is the average pooling layer, which is used to calculate the exposure level of the local area, E in is the enhanced image, N is the total number of pixels in the severe weather image, C is the total number of categories, is the unique hot encoding of the true value of the kth category, GT is the true value, is the probability that pixel n belongs to category k predicted by the image enhancement model, n∈[1,N], k∈[1,C], ω k Represents the weight parameter.

[0038] A severe weather image enhancement system based on physical prior knowledge, which inputs rainy / snowy images into a trained image enhancement model and outputs semantically segmented images;

[0039] The training process of the image enhancement model includes the training set construction module, the sharpening module, the image enhancement module, the semantic segmentation module and the loss construction module:

[0040] The training set construction module is used to obtain images of different driving environments in sunny and bad weather to construct a training data set;

[0041] The sharpening module is used to use the sunny image as a reference image, learn the distribution difference between the sunny image and the bad weather image, separate the weather component from the bad weather image, and generate a sharpened image;

[0042] The image enhancement module is used to enhance the clear image by introducing physical priors to obtain an enhanced image;

[0043] The semantic segmentation module is used to perform semantic segmentation on the enhanced image to obtain a semantic segmentation image with detailed features;

[0044] The loss building module is used to construct the total loss function and adjust the parameters of the image enhancement model.

[0045] Further, the sharpening module includes six sequentially connected convolutional layers;

[0046] The first three convolutional layers are used to process the severe weather image and downsample layer by layer to reduce the size of the severe weather image and extract low-level features of the severe weather image to obtain the first image;

[0047] The last three convolutional layers gradually upsample the first image through deconvolution operations, restore the first image to its original resolution, and ensure that the low-level features in the first image are preserved, thereby generating a sharpened image;

[0048] Batch regularization is applied after each convolutional layer to stabilize the feature distribution, and the ReLU activation function is used to enhance the nonlinear expression ability of the clarity module.

[0049] The advantages of the severe weather image enhancement method and system provided by the present invention based on physical prior knowledge are: combining prior knowledge and self-supervised learning strategies, and by designing multiple loss functions such as clarity, enhancement and semantic segmentation, the purpose of efficiently removing interference, enhancing clarity and improving semantic segmentation effects of rainy and snowy weather images is achieved. The system can separate weather components from images, generate high-quality images that meet actual application requirements under different weather conditions, and optimize visual perception effects through adaptive training, ultimately improving the safety and reliability of the autonomous driving system in severe weather such as rainy and snowy weather. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0051] Below, the technical solution of the present invention is described in detail through specific embodiments. Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific implementation disclosed below.

[0052] like Figure 1 As shown, the present invention proposes a severe weather image enhancement method and system based on physical prior knowledge, which inputs the image to be enhanced into a trained image enhancement model and outputs a semantic segmentation image.

[0053] This embodiment can improve the visual clarity and semantic segmentation performance of images under complex weather conditions. This method relies on physical prior knowledge and achieves separation and clarity of components through self-supervised learning. This embodiment uses rainy and snowy weather conditions as a type of severe weather to illustrate. The training process of the image enhancement model is as follows:

[0054] Step 1: Obtain images of different driving environments in sunny and bad weather to build a training data set;

[0055] Data collection: Use visual sensors to collect driving environment image data under sunny and rainy and snowy weather conditions. Sunny image I ref As a reference image, it does not contain rain or snow interference and is used for self-supervised learning of subsequent image enhancement models.

[0056] Step 2: Using the sunny image as a reference image, the clearing module learns the distribution difference between the sunny image and the bad weather image, separates the weather component from the bad weather image, and generates a cleared image;

[0057] The reference image is used to generate a sharpened image. The sharpening module learns the distribution difference between the sunny image and the bad weather image, thereby removing interference from the input rain and snow image under self-supervision, where the distribution difference specifically includes (a1) to (a3):

[0058] (a1) Learning the distribution differences of rain / snow weather components through the clarity module.

[0059] Specifically, by comparing sunny and rainy / snowy weather images, we learn the distribution characteristics of raindrops, water marks, snowflakes and other weather components in the image. The network uses this distribution difference to identify and separate the rain / snow components in the image, thereby generating a clear image without weather interference.

[0060] (a2) Learning the distribution differences of brightness and hue of rain / snow weather components through the sharpening module;

[0061] Rainy / snowy weather can cause the overall brightness of an image to decrease and the hue to shift. By learning the distribution difference, the system can identify these characteristic shifts in brightness and hue, thereby restoring the normal brightness and hue of the image after removing the weather component, making the generated image closer to the visual effect under sunny conditions.

[0062] (a3) Learning the distribution difference of local contrast of rain / snow weather components through the sharpening module;

[0063] The rain / snow component affects the contrast of local areas of the image, especially in areas covered by raindrops or snowflakes, where the contrast is low. The model restores the local contrast in the clear image by learning the distribution difference of local contrast, thereby improving the visual clarity of the image.

[0064] The sharpening module of this embodiment has six sequentially connected convolutional layers, the first three convolutional layers are used to process the severe weather image and downsample layer by layer to reduce the size of the severe weather image and extract the low-level features of the severe weather image to obtain the first image. The last three convolutional layers are gradually upsampled through deconvolution operations to restore the first image to the original resolution and ensure that the low-level features in the first image are retained, thereby generating a sharpened image.

[0065] In these six convolutional layers, the ResNetBlock structure is used to enhance the expressive power of the sharpening module. The residual connection of the two convolutional layers is used to achieve effective propagation and enhancement of feature information, ensuring the retention of image details and efficient processing.

[0066] In the clarity module, Batch regularization is applied after each convolution layer to stabilize the feature distribution, and the ReLU activation function is used to enhance the nonlinear expression ability of the network to improve the processing efficiency and stability of the system under complex weather conditions.

[0067] The setting of convolution kernel in the sharpening module: 7×7 convolution kernel is used in the first and last convolution layers to expand the receptive field, thereby enhancing the expressiveness of initial feature extraction and final image reconstruction; 3×3 convolution kernel is used in the remaining layers to ensure computational efficiency and accuracy of feature extraction. The final output of the sharpening module is a three-channel RGB image, that is, a sharpened image after removing rain / snow interference, which is used for subsequent semantic segmentation or other visual tasks.

[0068] Step 3: Introduce physical priors into the image enhancement module, perform image enhancement processing on the cleared image, and obtain an enhanced image;

[0069] Based on physical priors, the image enhancement method model makes the rain and snow image enhancement system more accurate in processing severe weather images, and the generated images are of higher quality. Specifically, a model is established based on atmospheric scattering theory to describe the scattering process of light in the atmosphere under rain and snow weather conditions. The model is based on the scattering and absorption effects of light when it propagates in the atmosphere, and is mainly used to estimate image degradation caused by haze, rain and snow weather, and effectively reduce the blur caused by light scattering, thereby improving the contrast and clarity of the image.

[0070] The atmospheric scattering theory model is as follows:

[0071] I(x)=J(x)t(x)+A(1-y(x));

[0072] Among them, I(x) is the pixel value of the severe weather image at position x, J(x) is the pixel value of the real image of the scene at position x, A represents the background brightness, and t(x) is the transmittance, which is defined as the light transmission rate from the scene to the camera.

[0073] By using gamma correction technology, the brightness and contrast of the image are adjusted to compensate for the brightness reduction and contrast shift caused by rainy and snowy weather;

[0074]

[0075] Among them, I in is the pixel value of the input severe weather image (usually in the range of 0 to 1), I out is the pixel value of the enhanced image, γ is the gamma value, when γ<1, it is used to enhance the dark details, and when γ>1, it is used to compress the bright details.

[0076] By adjusting the gamma value of the sharpened image, the visual effect of the sharpened image is improved, the image details are made clearer, and the expressiveness of the dark area is enhanced to ensure visual consistency under different lighting conditions.

[0077] Specifically, a parameter prediction module is implemented through a convolutional neural network (CNN), which is designed to be resolution-independent, differentiable, and understandable to ensure the versatility and stability of the image enhancement model under different input image resolutions.

[0078] The image enhancement module includes six convolutional layers connected in sequence. The convolution operation of each convolutional layer gradually doubles the number of channels and reduces the spatial size to half of the previous convolutional layer to extract the multi-scale feature information of the clear image step by step. The output of the last convolutional layer of the image enhancement module generates the parameters required for image enhancement; the extracted clear image is enhanced based on the required parameters.

[0079] For the i-th convolutional layer, its operation is implemented by the following formula:

[0080] X i =IN(σ(W i X i-1 +b i ));

[0081] Among them, W i 、b i is the weight of the image enhancement module, X i represents the output features of the i-th layer, σ is the LeakyReLU activation function, which is used to enhance the nonlinear ability of feature expression, and IN represents instance normalization, which is used to reduce the distribution deviation between batches, thereby improving the robustness of the model under different weather conditions.

[0082] Through the above structural design, the image enhancement module uses the convolutional layer to gradually extract features related to weather components during the forward propagation process, and maps them into specific image enhancement parameters, thereby significantly improving the visual effects under rainy / snowy weather conditions.

[0083] Step 4: Perform semantic segmentation on the enhanced image to obtain a semantic segmentation image with detailed features;

[0084] Downstream tasks: In autonomous driving systems, the main goal of image enhancement is to provide clearer and more accurate visual data for downstream detection and perception tasks to improve the robustness and safety of the system in complex environments. Through image enhancement, the system can restore the color and details of objects, so that the object attribute recognition model can more accurately analyze and judge the detailed features of objects. This helps the autonomous driving system make more accurate decisions in complex scenarios, such as making avoidance or overtaking decisions based on vehicle type and direction.

[0085] Step 5: Construct a total loss function and adjust the parameters of the image enhancement model

[0086] Design a total loss function to implement system training and parameter updates, so that the system output meets the expected effect and improves the performance of downstream tasks. The total loss function specifically includes clarity loss, exposure loss, and semantic segmentation loss.

[0087] Clarification loss: This loss term is used to measure the difference between the generated clear image and the target clear image to ensure that the rain / snow components are effectively removed. By minimizing the clarity loss, the image enhancement model can enhance the effect of removing rain and snow interference during training and improve the clarity of the image.

[0088] Exposure loss: This loss term is used to adjust the overall brightness level of the image to correct the brightness deviation caused by rainy and snowy weather. By optimizing the exposure loss, the image enhancement model can dynamically adjust the brightness of the image so that the generated enhanced image has appropriate brightness performance and is more in line with the visual characteristics under sunny conditions.

[0089] Semantic segmentation loss: This loss term is used to ensure the performance of the enhanced image in the semantic segmentation task. By introducing semantic segmentation loss, the image enhancement model retains the semantic information in the image during the sharpening process, thereby maintaining the object boundaries and semantic features in the processed image, improving the detection and recognition accuracy in the intelligent driving system.

[0090] By using a reference image I taken in clear weather ref , define the prior loss function, prior loss L ref for:

[0091]

[0092] By learning,the prior loss term, the image enhancement model learns the distribution,differences of clear weather and rainy and snowy weather characteristics to,improve the recognition and separation capabilities of severe weather,images.

[0093] The structural similarity index (SSIM) is used to evaluate the sharpened image C output by the sharpening module. in With the input severe weather image I in and reference image I ref Similarity, structural similarity index loss L ssim for:

[0094]

[0095] In order to ensure that the enhanced image E in The brightness and contrast are in line with expectations, and exposure loss is introduced to quantify the deviation between the exposure level of the local area and the ideal exposure level. The exposure loss L exp for:

[0096]

[0097] The weighted cross entropy loss is used to quantify the difference between the segmentation map predicted by the model and the true label, and the semantic segmentation loss L seg for:

[0098]

[0099] Among them, C(I ref ) is the cleared image output by the clearing module, is the all-0 feature map, ∥·∥2 is the binary norm, SSIM is the structural index, C in is the cleared image output by the clearing module, I in For severe weather images, I ref is the reference image, s is a set constant, which is set according to the source of the reference image to measure the similarity between the reference image and the severe weather image, ∥·∥1 is a norm, M is the number of local areas divided by the enhanced image, is an average pooling layer, which is used to calculate the exposure level of the local area, where the size of the local area is set to 16×16 pixels, E in is the enhanced image, N is the total number of pixels in the severe weather image, C is the total number of categories, is the unique hot encoding of the true value of the kth category, GT is the true value, is the probability that pixel n belongs to category k predicted by the image enhancement model, n∈[1,N], k∈[1,C], ω k Represents the weight parameter.

[0100] The semantic segmentation loss term is used to ensure that the enhanced image retains the boundary and semantic information in the semantic segmentation task.

[0101] This embodiment combines prior knowledge and self-supervised learning strategies, and achieves the purpose of efficiently removing interference, enhancing clarity, and improving semantic segmentation effects for rainy and snowy weather images by designing multiple loss functions such as clarity, enhancement, and semantic segmentation. The system can separate weather components from images, generate high-quality images that meet actual application needs under different weather conditions, and optimize visual perception effects through adaptive training, ultimately improving the safety and reliability of the autonomous driving system in severe weather such as rain and snow.

[0102] That is, this embodiment has the following advantages:

[0103] (1) Introducing physical prior knowledge to improve image enhancement: This embodiment uses physical prior knowledge such as atmospheric scattering theory and gamma correction, so that the system can more accurately estimate the impact of haze and light scattering on image clarity when processing rainy and snowy weather images, thereby effectively reducing image blur and contrast loss caused by these factors. Compared with traditional image denoising or filtering methods, this embodiment not only restores the brightness and contrast of the image, but also retains more image details, thereby significantly improving the visual quality of the image.

[0104] (2) Enhance semantic segmentation performance and improve autonomous driving safety: This embodiment combines semantic segmentation loss to ensure that image enhancement not only improves visual clarity, but also ensures the image's performance in semantic segmentation tasks. In an autonomous driving system, the accuracy of object detection and semantic segmentation is directly related to driving safety, and the model of this embodiment can enhance the semantic information of the image under adverse weather conditions, thereby improving the accuracy of target detection and recognition by the autonomous driving system, and enhancing the reliability and safety of the autonomous driving system in rainy and snowy weather.

[0105] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method and system for enhancing severe weather images based on physical prior knowledge, characterized in that: Input the image to be enhanced into the trained image enhancement model and output the semantic segmentation image; The training process of the image enhancement model is as follows: Obtain images of different driving environments in sunny and bad weather to build a training data set; Using the sunny image as the reference image, the clearing module learns the distribution difference between the sunny image and the bad weather image, separates the weather component from the bad weather image, and generates a clear image. Introducing physical priors into the image enhancement module, the image enhancement processing is performed on the cleared image to obtain the enhanced image; Perform semantic segmentation on the enhanced image to obtain a semantic segmentation image with detailed features; Construct a total loss function and adjust the parameters of the image enhancement model.

2. The severe weather image enhancement method based on physical prior knowledge according to claim 1 is characterized in that: The sharpening module includes six convolutional layers connected in sequence; The first three convolutional layers are used to process the severe weather image and downsample layer by layer to reduce the size of the severe weather image and extract low-level features of the severe weather image to obtain the first image; The last three convolutional layers gradually upsample the first image through deconvolution operations, restore the first image to its original resolution, and ensure that the low-level features in the first image are preserved, thereby generating a sharpened image; Batch regularization is applied after each convolutional layer to stabilize the feature distribution, and the ReLU activation function is used to enhance the nonlinear expression ability of the clarity module.

3. The severe weather image enhancement method based on physical prior knowledge according to claim 1 is characterized in that: The image enhancement module includes six convolutional layers connected in sequence; The convolution operation of each convolution layer gradually doubles the number of channels and reduces the spatial size to half of the previous convolution layer to extract the multi-scale feature information of the clear image step by step. The output of the last convolution layer of the image enhancement module generates the parameters required for image enhancement; the extracted clear image is enhanced based on the required parameters.

4. The severe weather image enhancement method based on physical prior knowledge according to claim 1, characterized in that: Physical priors are introduced into the image enhancement module to perform image enhancement processing on the cleared image, specifically: A model was established based on atmospheric scattering theory to describe the scattering process of light in the atmosphere under severe weather conditions, and the effects of haze and suspended water molecules on the clear image were obtained; Gamma correction technology is used to adjust the brightness and contrast of the cleared image to compensate for the brightness reduction and contrast shift caused by bad weather, thereby obtaining an enhanced image.

5. The severe weather image enhancement method based on physical prior knowledge according to claim 4 is characterized in that: The atmospheric scattering theory model is as follows: I(x)=J(x)t(x)+A(1-t(x)); Among them, I(x) is the pixel value of the severe weather image at position x, J(x) is the pixel value of the real image of the scene at position x, A represents the background brightness, and t(x) is the transmittance, which is defined as the light transmission rate from the scene to the camera.

6. The severe weather image enhancement method based on physical prior knowledge according to claim 4 is characterized in that: Gamma correction technology is used to adjust the brightness and contrast of the clear image, specifically: Among them, I in is the pixel value of the input severe weather image, I out is the pixel value of the enhanced image, γ is the gamma value, when γ<1, it is used to enhance the dark details, and when γ>1, it is used to compress the bright details.

7. The severe weather image enhancement method based on physical prior knowledge according to claim 1, characterized in that: The total loss function includes sharpening loss, exposure loss and semantic segmentation loss.

8. The severe weather image enhancement method based on physical prior knowledge according to claim 7 is characterized in that: The clarity loss includes a priori loss and a structural similarity index loss; The prior loss L ref for: The structural similarity index loss L ssim for: The exposure loss L exp for: The semantic segmentation loss L seg for: Among them, C(I ref ) is the cleared image output by the clearing module, is the all-0 feature map, ∥·∥2 is the binary norm, SSIM is the structural index, C in is the cleared image output by the clearing module, I in For severe weather images, I ref is the reference image, s is a set constant to measure the similarity between the reference image and the severe weather image, ∥·∥1 is a norm, M is the number of local regions divided by the enhanced image, is the average pooling layer, which is used to calculate the exposure level of the local area, E in is the enhanced image, N is the total number of pixels in the severe weather image, C is the total number of categories, is the unique hot encoding of the true value of the kth category, GT is the true value, is the probability that pixel n belongs to category k predicted by the image enhancement model, n∈[1,N], k∈[1,C], ω k Represents the weight parameter.

9. A severe weather image enhancement system based on physical prior knowledge, characterized in that: Input the rainy / snowy image into the trained image enhancement model and output the semantic segmentation image; The training process of the image enhancement model includes the training set construction module, the sharpening module, the image enhancement module, the semantic segmentation module and the loss construction module: The training set construction module is used to obtain images of different driving environments in sunny and bad weather to construct a training data set; The sharpening module is used to use the sunny image as a reference image, learn the distribution difference between the sunny image and the bad weather image, separate the weather component from the bad weather image, and generate a sharpened image; The image enhancement module is used to enhance the clear image by introducing physical priors to obtain an enhanced image; The semantic segmentation module is used to perform semantic segmentation on the enhanced image to obtain a semantic segmentation image with detailed features; The loss building module is used to construct the total loss function and adjust the parameters of the image enhancement model.

10. The severe weather image enhancement system based on physical prior knowledge according to claim 9, characterized in that: The sharpening module includes six convolutional layers connected in sequence; The first three convolutional layers are used to process the severe weather image and downsample layer by layer to reduce the size of the severe weather image and extract low-level features of the severe weather image to obtain the first image; The last three convolutional layers gradually upsample the first image through deconvolution operations, restore the first image to its original resolution, and ensure that the low-level features in the first image are preserved, thereby generating a sharpened image; Batch regularization is applied after each convolutional layer to stabilize the feature distribution, and the ReLU activation function is used to enhance the nonlinear expression ability of the clarity module.

Images