An AI Image Processing Method for Electronic Rearview Mirrors

Abstract

This invention relates to the field of image processing technology, specifically to an AI image processing method for electronic rearview mirrors. In this invention, an original rearview mirror image is acquired and decoupled analysis is performed based on a physical degradation model to generate a color attenuation component and a detail interference component. The color attenuation component is processed by a global color and illumination reconstruction network, and a scene semantic attention map is generated using an embedded scene semantic attention module to guide adaptive reconstruction, outputting a first optimized image. Simultaneously, the detail interference component and the scene semantic attention map are input into a local structure to guide a detail repair network, and differential adjustment is performed to output a second optimized image. Pixel-level fusion weights are generated using the scene semantic attention map as a guide, and the two optimized images are weighted and fused to output the final optimized image. This method achieves adaptive and differential intensity control of different semantic regions through the scene semantic attention map, solving the problem of inappropriate regional effects caused by uniform processing of the entire image.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to an AI image processing method for electronic rearview mirrors. Background Technology

[0002] With the rapid development of intelligent driving technology, electronic rearview mirrors, as a crucial sensing component replacing traditional optical lenses, directly impact driving safety and the reliability of decision support through their image quality. In complex and ever-changing driving environments, such as fog, rain, nighttime conditions, or sudden changes in brightness inside and outside tunnels, rearview mirror images captured by in-vehicle cameras are prone to severe quality degradation. This degradation is not caused by a single factor, but rather by the combined effects of uneven lighting, atmospheric scattering leading to overall color and contrast attenuation, and the loss of local details caused by noise, motion blur, and water droplet interference. Most existing image enhancement or restoration methods employ a unified global processing framework, such as contrast stretching based on histogram equalization or end-to-end mapping networks based on deep learning. While these methods can improve the overall visual effect, they neglect the spatial heterogeneity of the image content itself.

[0003] Specifically, a typical driving image contains multiple semantic regions, including the sky, roads, vehicles, vegetation, and buildings. These different regions experience significantly different types and degrees of degradation during the physical imaging process. For example, the sky is primarily affected by atmospheric light, resulting in a washed-out appearance with minimal detail loss; while vehicle and pedestrian areas are more susceptible to noise and blurring. Current technologies apply the same enhancement or restoration intensity to the entire image, lacking the ability to differentiate between semantic regions. This leads to situations where some areas in the processed image may suffer from color distortion or noise amplification due to over-enhancement, while other areas may be under-processed, leaving details blurred. This "one-size-fits-all" approach makes it difficult to achieve an ideal balance between global color naturalness and local detail clarity, limiting the ability of electronic rearview mirrors to provide stable, clear, and realistic scene perception in complex environments. Summary of the Invention

[0004] The purpose of this invention is to provide an AI image processing method for electronic rearview mirrors to solve the problems mentioned in the background art. Specific technical problems include how to adaptively and differentially adjust the intensity and fuse information for the two key processing stages—global color illumination correction and local detail restoration—based on the prior characteristics of different semantic regions in the image. This addresses the problems of inappropriate processing effects in certain areas, inconsistent overall image optimization results, and difficulty in balancing naturalness and clarity caused by the uniform processing of the entire image in existing methods.

[0005] To achieve the above objectives, the present invention aims to provide an AI image processing method for electronic rearview mirrors, specifically including the following method steps: S1. Acquire raw rearview mirror images under different environments, including rain, fog, snow, dust storms, and haze. By acquiring raw rearview mirror images covering various harsh environments such as rain, fog, snow, dust storms, and haze, a comprehensive data foundation is provided for subsequent processing. This step ensures that the method can learn and adapt to the complex relationship between environmental degradation and image semantic content, laying the groundwork for data diversity for subsequent differentiated processing based on the prior characteristics of different semantic regions in the image, thereby resolving the problem of inconsistent overall image optimization results.

[0006] S2. Decoupled analysis based on a physical degradation model is performed on the original rearview mirror image to generate color attenuation and detail interference components. Using dark channel priors and atmospheric illumination estimation methods, the atmospheric light component and non-uniform transmittance distribution in the image are calculated and separated. This leads to the derivation of the pure scene reflected light component with the given color and illumination conditions. The difference between the pure scene reflected light component and the original rearview mirror image is quantified as a color attenuation component. The degradation component dominated by the color attenuation component is subtracted from the original rearview mirror image, and the remaining part is defined as the detail interference component. The construction process of the physical degradation model specifically includes: Based on the physical mechanism of light propagation under different environments, it is determined that the degradation of the original rearview mirror image consists of two parts: illumination color degradation and structural detail degradation. These two parts of degradation are then modeled in a linear superposition manner to construct a physical degradation model.

[0007] Step S2 decouples the original rearview mirror image based on a physical degradation model, separating the mixed degradation into color attenuation components and detail interference components. The purpose of this step is to explicitly separate the core degradation factors targeted by the two tasks of global color illumination correction and local detail restoration from the physical representation level. This provides a clear input for the subsequent targeted and independently controllable processing of the two specialized networks, avoiding improper processing results caused by mutual interference between different degradations. It is a key preparatory step for achieving differentiated intensity control and information fusion.

[0008] S3. The color attenuation component is input into the global color and illumination reconstruction network for processing. The global color and illumination reconstruction network generates a scene semantic attention map through an embedded scene semantic attention module, specifically including: The global color and illumination reconstruction network is a deep convolutional neural network based on an encoder-decoder structure; A parallel semantic segmentation branch is used to parse the deep feature map, identify and label regions of different semantic categories, and calculate an attention weight matrix based on the prior knowledge of each semantic category. This attention weight matrix is ​​then used as the scene semantic attention map.

[0009] Adaptive reconstruction of the color attenuation component is performed based on the scene semantic attention map to generate the first optimized image, specifically including: During the upsampling and feature fusion process in the decoder, the scene semantic attention map is used as a spatial weight matrix and multiplied pixel by pixel with the intermediate feature map in the decoder for semantically guided feature modulation to generate the modulated feature map. Based on the modulated feature map, a global color and illumination correction map is learned and output. The global color and illumination correction map and the color attenuation component are added pixel by pixel, and the corrected result is used as the first optimized image.

[0010] Step S3 inputs the color attenuation component into the global color and illumination reconstruction network and uses the embedded scene semantic attention module to generate a scene semantic attention map. Based on the prior characteristics of different semantic regions, the scene semantic attention map serves as a spatial weight matrix to guide adaptive reconstruction during feature modulation, thereby performing differentiated intensity control on global color and illumination correction, solving the problem of color distortion or unnatural illumination in certain areas caused by the uniformization of the entire image, and improving the naturalness of the overall image.

[0011] S4. Input the detail interference components and the scene semantic attention map into the local structure-guided detail restoration network for processing. Based on the scene semantic attention map, differentiate the detail restoration intensity of different semantic regions and output the second optimized image, specifically including: Input the detail interference components and the scene semantic attention map into the local structure to guide the detail repair network; The local structure-guided detail restoration network uses the scene semantic attention map to dynamically generate convolution kernel parameters related to spatial location to extract differential features of detail interference components, and injects the scene semantic attention map into the feature channel through a conditional normalization layer. The processed features are upsampled and fused to output a second optimized image that improves local texture clarity, edge sharpness, and noise suppression.

[0012] Step S4 inputs the detail interference component and the scene semantic attention map into the local structure-guided detail restoration network, and differentiates the detail restoration intensity of different semantic regions based on the scene semantic attention map; by injecting semantic information through dynamic convolution kernels and conditional normalization, adaptive optimization of local texture clarity, edge sharpness and noise suppression is achieved, solving the problem of over-sharpening or blurring of certain areas caused by the uniformization of the whole image, and improving the clarity of the overall image.

[0013] S5. Guided by the scene semantic attention map, generate pixel-level fusion weights that are dynamically related to the semantic region, specifically including: The scene semantic attention map is input into a weight generation network consisting of consecutive convolutional layers. This weight generation network learns the transition and weight allocation relationships between and within different semantic category regions through convolutional operations, and outputs a single-channel weight map with the same size as the original input image as the pixel-level fusion weight. The specific calculation process of the weight value at each pixel position in the pixel-level fusion weight includes: The weight generation network comprehensively reads the semantic category encoding value of the scene semantic attention map at the pixel position, as well as the semantic category and attention intensity values ​​of all pixels in its surrounding local neighborhood, and performs non-linear combination and mapping to output a definite scalar between zero and one as the weight value of the pixel position.

[0014] The pixel information of the first and second optimized images is weighted and fused according to pixel-level fusion weights to generate and output the final optimized image. The weighted fusion process specifically includes: The first optimized image, the second optimized image, and the pixel-level fusion weight map are mapped one-to-one at their spatial pixel positions, and the fusion formula is used for calculation. The fusion formula multiplies the color value of each channel of each pixel in the first optimized image with the weight value of the corresponding point in the pixel-level fusion weight map to obtain intermediate result one; Multiply the color value of each channel of each pixel in the second optimized image with the complementary weight value obtained by subtracting the weight value of the corresponding point in the pixel-level fusion weight map to obtain the second intermediate result; The corresponding channel values ​​of intermediate result one and intermediate result two at the same pixel position are added together, and the sum is used as the color value of the final optimized image at that pixel position and in that channel.

[0015] Step S5 uses the scene semantic attention map as a guide to generate pixel-level fusion weights, and then performs weighted fusion on the first optimized image (global color illumination correction result) and the second optimized image (local detail restoration result) based on these weights. This adaptive information fusion based on the prior characteristics of semantic regions achieves differentiated coordination of the processing intensity of different semantic regions, solving the problems of inappropriate regional effects, inconsistent overall image optimization results, and difficulty in balancing naturalness and clarity caused by the uniform processing of the whole image in existing methods.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention introduces a scene semantic attention map and, based on the prior characteristics of different semantic regions in the image, adaptively and differentially adjusts the intensity and fuses information for global color illumination correction and local detail restoration. This effectively solves the problems of improper processing in certain areas, inconsistent overall image optimization results, and difficulty in balancing naturalness and clarity caused by the uniform processing of the entire image in existing methods. As a result, it significantly improves the overall visual quality and environmental adaptability of rearview mirror images. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall method steps of the present invention; Figure 2 This is the core flowchart of the application of scene semantic attention graph in this invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Next, please refer to Figure 1 The purpose of this embodiment is to provide an AI image processing method for electronic rearview mirrors, which includes the following steps: S1. Directly receive raw rearview mirror images captured and transmitted in real time by the image sensor mounted on a designated side of the vehicle under different environments, including but not limited to rain, fog, snow, sandstorms and haze. The acquired raw rearview mirror images are the initial inputs to be processed in subsequent steps.

[0020] S2. Perform decoupling analysis on the original rearview mirror image based on a physical degradation model to obtain the color attenuation component and detail interference component, specifically including: A physical degradation model is established for imaging characteristics under different environments. This model explicitly characterizes the degradation process of the original rearview mirror image as a linear superposition of illumination color degradation and structural detail degradation. The construction of the physical degradation model is based on a clear analysis of the physical mechanism of light propagation under different environments. It is determined that image degradation mainly consists of two parts: illumination color degradation and structural detail degradation. Illumination color degradation is caused by atmospheric light composition and non-uniform transmittance distribution due to suspended particles in the atmosphere, resulting in color shift, contrast reduction, and uneven illumination. Structural detail degradation is caused by random occlusion and scattering by particles, resulting in texture blurring, edge loss, and noise interference. These two parts of degradation are then modeled in a linear superposition manner, thus establishing a physical degradation model that explicitly characterizes the degradation process of the original rearview mirror image as a linear superposition of illumination color degradation and structural detail degradation.

[0021] Based on this physical degradation model, the first step in decoupling the analysis of the original rearview mirror image is to calculate and separate the atmospheric light component and non-uniform transmittance distribution caused by suspended particles in the atmosphere (such as fog droplets, rain, snow, and dust) through dark channel prior and atmospheric illumination estimation methods. This allows for the derivation of the color and illumination state that the pure scene reflected light component should possess. The difference between this part and the original rearview mirror image is precisely quantified as a color attenuation component, which reflects the overall color shift, contrast reduction, and uneven illumination phenomena in the image. The specific process of the dark channel prior and atmospheric illumination estimation methods is as follows: Based on the dark channel prior principle, the dark channel is located and determined in the original rearview mirror image, and the global atmospheric light composition is directly estimated from the brightest pixel region in the dark channel. Secondly, using the dark channel and the atmospheric light composition, a non-uniform transmittance distribution map describing the attenuation of light in the medium is calculated according to the physical formula of the atmospheric scattering model. Finally, the original rearview mirror image, the estimated atmospheric light composition, and the non-uniform transmittance distribution are substituted into the inverse transformation formula of the atmospheric scattering model for calculation. The output of this calculation result is the derived color and illumination state of the reflected light composition of a pure scene that is theoretically unaffected by atmospheric suspended particles.

[0022] The second step of the decoupling analysis is to substitute the obtained color attenuation component into the physical degradation model in reverse, thereby subtracting the degradation part dominated by color and illumination attenuation from the original image. The remaining part mainly includes texture blurring caused by random occlusion of particles and scattering, edge loss, and random noise introduced by the imaging sensor. This part is defined as the detail interference component, which focuses on the loss and interference of local structural information of the image.

[0023] Finally, through the above decoupling analysis process, the original rearview mirror image is clearly decomposed into two independent parts: a color attenuation component and a detail interference component, for subsequent processing by the network.

[0024] Please see Figure 2 S3. The color attenuation component is input into the global color and illumination reconstruction network for processing. The global color and illumination reconstruction network generates a scene semantic attention map through the embedded scene semantic attention module, specifically including: The color attenuation component is input into a global color and illumination reconstruction network for processing. This global color and illumination reconstruction network is a deep convolutional neural network based on an encoder and decoder structure. The encoder part extracts multi-level features from the input color attenuation component through stacked convolutional layers and downsampling layers to obtain deep feature maps at different scales. Based on the deep features extracted by the encoder, the global color and illumination reconstruction network generates a scene semantic attention map through its embedded scene semantic attention module. Specifically, a parallel semantic segmentation branch is first used to parse the deep feature map, identify and label regions representing different semantic categories such as sky, road, vehicle, building and vegetation. Then, the scene semantic attention module calculates an attention weight matrix based on the prior knowledge of each semantic category (e.g., the sky region usually has severe illumination attenuation, while the vehicle region needs to maintain its inherent color attributes). This attention weight matrix is ​​the scene semantic attention map, where the weight value of each pixel position clearly represents the intensity of attention and correction required for the semantic category to which that position belongs in the global color and illumination reconstruction process.

[0025] After generating the scene semantic attention map, the global color and illumination reconstruction network adaptively reconstructs the color attenuation component based on the scene semantic attention map, specifically including: In the decoder part of the global color and illumination reconstruction network, during the upsampling and feature fusion of its deep feature maps, the scene semantic attention map is used as a spatial weight matrix and multiplied pixel by pixel with the intermediate feature map in the decoder. This allows for differentiated enhancement or suppression of feature responses in different semantic regions, achieving semantically guided feature modulation of the color attenuation component. The modulated feature map is then processed by subsequent convolutional layers, which learn and output a global color and illumination correction map. The global color and illumination correction map is then added pixel by pixel to the color attenuation component, ultimately generating the first optimized image with improved global color consistency, illumination uniformity, and contrast.

[0026] S4. Input the detail interference components and the scene semantic attention map into the local structure-guided detail restoration network for processing. Based on the scene semantic attention map, differentiate the detail restoration intensity of different semantic regions and output the second optimized image, specifically including: The detail interference components and the scene semantic attention map are input into a local structure-guided detail restoration network for processing. This network is a multi-layered deep neural network. The detail restoration intensity for different semantic regions is differentially adjusted based on the scene semantic attention map, specifically including: Based on the attention weights corresponding to different semantic regions in the scene semantic attention map, a set of convolution kernel parameters related to spatial location is dynamically generated. These convolution kernel parameters are used to construct dynamic convolutional layers, enabling the local structure-guided detail restoration network to adaptively apply different intensities of feature extraction and enhancement operations to the detail interference components of different regions (such as vehicle edge regions that require fine texture restoration and sky regions that can be accepted for smoothing). The local structure-guided detail restoration network injects semantic and intensity information encoded by the scene semantic attention map into the feature channel through the conditional normalization layer, thereby guiding the network to perform targeted restoration of details lost and noise interference in key areas at the feature level. The feature maps that have undergone the above differential regulation are progressively upsampled and feature fused in the network decoder, and finally output a second optimized image that is optimized in terms of local texture clarity, edge sharpness and noise suppression.

[0027] S5. Guided by the scene semantic attention map, generate pixel-level fusion weights that are dynamically related to the semantic region, specifically including: The scene semantic attention map is input into a lightweight weight generation network, which consists of several consecutive convolutional layers. This network receives the scene semantic attention map as its sole input and learns and models the transition and weight allocation relationships between and within different semantic category regions during image fusion through convolutional operations. The specific convolutional operation process is as follows: In the weight generation network, each pre-defined convolutional layer uses a set of fixed-size convolutional kernels with defined parameters to process the data input to that layer. During the operation, each convolutional kernel acts as a sliding window, traversing all spatial locations of the input data. At each current location, the value within the local area covered by the convolutional kernel is multiplied element-wise with the kernel's own parameters. All products are then summed, and a fixed bias parameter corresponding to that convolutional kernel is added to generate a new value for the output data at that location. This process is completed by sliding across the entire input data, thus generating the output feature map of that convolutional layer. The weight generation network, through multiple consecutive such convolutional layers, processes the input scene... The semantic attention map sequentially performs the multiplication, accumulation, and sliding calculations determined above, progressively extracting and transforming features, and finally outputting a single-channel weight map with the exact same size as the original input image, i.e., pixel-level fusion weights. The weight value at each pixel position is a definite scalar between zero and one, its magnitude dynamically determined by the semantic category and attention intensity of the input scene semantic attention map at the corresponding position and surrounding area. This ensures that in areas requiring emphasis on color and lighting reconstruction effects, the weights tend to favor the first optimized image, while in areas requiring enhanced details and structure, the weights tend to favor the second optimized image, achieving a smooth and natural transition of weights at the boundaries of different semantic regions. The specific calculation of the weight values ​​is as follows: The weight generation network extracts and transforms features from the input scene semantic attention map layer by layer through its continuous convolutional layers. During the learning process, for each specific pixel position on the final single-channel weight map, the convolution operation of the weight generation network comprehensively reads the semantic category encoding value of the input scene semantic attention map at the corresponding pixel position, as well as the semantic category and attention intensity values ​​of all pixels in a local neighborhood around that position. By performing non-linear combination and mapping of this information, the activation function at the end of the network outputs a definite scalar between zero and one, which is the final weight value of that pixel position.

[0028] The pixel information of the first and second optimized images is weighted and fused according to pixel-level fusion weights. The specific process is as follows: The first optimized image, the second optimized image, and the pixel-level fusion weight map are mapped one-to-one at their spatial pixel locations. A fusion formula is used to calculate the weighted intermediate result 1 by multiplying the color value of each channel of each pixel in the first optimized image with the weight value of the corresponding point in the pixel-level fusion weight map. Simultaneously, the color value of each channel of each pixel in the second optimized image is multiplied with the complementary weight value obtained by subtracting the weight value of the corresponding point in the pixel-level fusion weight map from the first formula, resulting in a weighted intermediate result 2. Subsequently, the corresponding channel values ​​of intermediate result 1 and intermediate result 2 at the same pixel location are added together, and the sum is the color value of the final optimized image at that pixel location and in that channel. This calculation is performed by traversing all pixel locations of the entire image to generate and output a final optimized image that integrates the color and lighting advantages of the first optimized image and the detail and structure advantages of the second optimized image at all pixels.

[0029] As can be seen from the above description, the AI ​​image processing method for electronic rearview mirrors provided in this embodiment has the following technical effects: The original rearview mirror image is decoupled into color attenuation and detail interference components using a physical degradation model. These components are then input into a global color and illumination reconstruction network and a local structure-guided detail restoration network, respectively. The two networks share a scene semantic attention map, and based on the prior characteristics of this map for different semantic regions, adaptive and differentiated intensity adjustments are made to global color and illumination correction and local detail restoration. Finally, pixel-level fusion weights are generated using the same scene semantic attention map as a guide, and the optimization results from the two stages are weighted and fused. This effectively solves the problems of improper processing in certain areas, inconsistent overall image optimization results, and difficulty in balancing naturalness and sharpness caused by the uniformization of the entire image in existing methods.

[0030] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI image processing method for an electronic rearview mirror, characterized in that, The methods and steps include the following: S1. Acquire raw rearview mirror images collected under different environments; S2. Perform decoupling analysis on the original rearview mirror image based on the physical degradation model to generate color attenuation components and detail interference components; S3. The color attenuation component is input into the global color and illumination reconstruction network for processing. The global color and illumination reconstruction network generates a scene semantic attention map through an embedded scene semantic attention module, and performs adaptive reconstruction of the color attenuation component based on the scene semantic attention map to generate the first optimized image. S4. Input the detail interference component and the scene semantic attention map into the local structure-guided detail repair network for processing, and perform differentiated adjustment of the detail repair intensity of different semantic regions according to the scene semantic attention map to output the second optimized image; S5. Guided by the scene semantic attention map, generate pixel-level fusion weights that are dynamically related to the semantic region, and perform weighted fusion of the pixel information of the first optimized image and the second optimized image according to the pixel-level fusion weights to generate and output the final optimized image.

2. The AI ​​image processing method for electronic rearview mirrors according to claim 1, characterized in that, The construction process of the physical degradation model specifically includes: Based on the physical mechanism of light propagation under different environments, it is determined that the degradation of the original rearview mirror image consists of two parts: illumination color degradation and structural detail degradation. These two parts of degradation are then modeled in a linear superposition manner to construct a physical degradation model.

3. The AI ​​image processing method for electronic rearview mirrors according to claim 1, characterized in that, The process of generating the color attenuation component specifically includes: By using dark channel prior and atmospheric illumination estimation methods, the atmospheric light component and non-uniform transmittance distribution in the image are calculated and separated. Then, the pure scene reflected light component with color and illumination status is derived, and the difference between the pure scene reflected light component and the original rearview mirror image is quantified as a color attenuation component.

4. The AI ​​image processing method for electronic rearview mirrors according to claim 1, characterized in that, The generation process of the detailed interference components specifically includes: The degradation component dominated by color attenuation is subtracted from the original rearview mirror image, and the remaining part is defined as the detail interference component.

5. The AI ​​image processing method for electronic rearview mirrors according to claim 1, characterized in that, The process of generating the scene semantic attention map specifically includes: The global color and illumination reconstruction network is a deep convolutional neural network based on an encoder-decoder structure; A parallel semantic segmentation branch is used to parse the deep feature map, identify and label regions of different semantic categories, and calculate an attention weight matrix based on the prior knowledge of each semantic category. This attention weight matrix is ​​then used as the scene semantic attention map.

6. The AI ​​image processing method for electronic rearview mirrors according to claim 5, characterized in that, The process of generating the first optimized image specifically includes: During the upsampling and feature fusion process in the decoder, the scene semantic attention map is used as a spatial weight matrix and multiplied pixel by pixel with the intermediate feature map in the decoder for semantically guided feature modulation to generate the modulated feature map. Based on the modulated feature map, a global color and illumination correction map is learned and output. The global color and illumination correction map is added pixel by pixel to the color attenuation component, and the corrected result is used as the first optimized image.

7. The AI ​​image processing method for electronic rearview mirrors according to claim 1, characterized in that, The process of generating the second optimized image specifically includes: The detailed interference components and the scene semantic attention map are input into the local structure-guided detail repair network; The local structure-guided detail repair network extracts differential features from detail interference components by dynamically generating convolutional kernel parameters related to spatial location based on the scene semantic attention map, and injects the scene semantic attention map into the feature channel through a conditional normalization layer. The processed features are upsampled and fused to output a second optimized image that improves local texture clarity, edge sharpness, and noise suppression.

8. The AI ​​image processing method for electronic rearview mirrors according to claim 1, characterized in that, The generation process of the pixel-level fusion weights specifically includes: The scene semantic attention map is input into a weight generation network consisting of consecutive convolutional layers. This weight generation network learns the transition and weight distribution relationships between and within different semantic category regions through convolutional operations, and outputs a single-channel weight map with the same size as the original input image as pixel-level fusion weights.

9. The AI ​​image processing method for electronic rearview mirrors according to claim 8, characterized in that, The calculation process for the weight value at each pixel position in the pixel-level fusion weight specifically includes: The weight generation network comprehensively reads the semantic category encoding value of the scene semantic attention map at the pixel position, as well as the semantic category and attention intensity values ​​of all pixels in its surrounding local neighborhood, and performs nonlinear combination and mapping to output a definite scalar between zero and one as the weight value of the pixel position.

10. The AI ​​image processing method for electronic rearview mirrors according to claim 1, characterized in that, The weighted fusion process specifically includes: The first optimized image, the second optimized image, and the pixel-level fusion weight map are mapped one-to-one at their spatial pixel positions, and the fusion formula is used for calculation. The fusion formula multiplies the color value of each channel of each pixel in the first optimized image with the weight value of the corresponding position point in the pixel-level fusion weight map to obtain intermediate result one; Multiply the color value of each channel of each pixel in the second optimized image with the complementary weight value obtained by subtracting the weight value of the corresponding point in the pixel-level fusion weight map to obtain the second intermediate result; The corresponding channel values ​​of intermediate result one and intermediate result two at the same pixel position are added together, and the sum is used as the color value of the final optimized image at that pixel position and in that channel.

Images