CMS image glare suppression method

Abstract

The invention belongs to the field of computer vision and image processing, and discloses a CMS image glare suppression method, which comprises the following steps: S1, establishing a learning sample for CMS image glare suppression; s2, building and training of an intelligent model for CMS image glare suppression are realized; and S3, glare suppression of the unknown CMS image. According to the invention, the attention module is added to guide the model to focus on the glare area in the image, and the multi-scale module is added to improve the ability of the model to capture glare of different scales, so that the glare suppression effect of the intelligent model is improved, and better CMS image high-quality imaging is realized.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and image processing, and in particular to a method for suppressing glare in CMS images. Background Technology

[0002] Rearview mirrors are the primary way for drivers to obtain information about the outside of their vehicle during a journey. However, traditional glass rearview mirrors suffer from problems such as high wind resistance (noise), narrow field of view, susceptibility to rain and snow, and inability to integrate with intelligent driving systems. Therefore, electronic rearview mirrors are expected to gradually replace traditional glass rearview mirrors in the field of intelligent driving and become a future trend. Electronic rearview mirrors are systems composed of cameras and monitors, referred to as "camera-monitor system" (CMS) in the new GB15084 standard. During driving, the CMS captures images of the road, vehicles, pedestrians, etc., to the sides and rear of the vehicle, processes and synthesizes them through software, and finally transmits them to the LCD display inside the vehicle, providing the driver with a good road condition image. Due to the variability of driving scenarios, such as scene type, light intensity, light distribution, and road conditions, achieving high-quality imaging with the CMS is extremely difficult, with glare caused by relatively strong light being the most prominent issue. Summary of the Invention

[0003] To address this, this invention proposes a glare suppression method for CMS images. Considering that the pixel proportion of glare in image deglare tasks is relatively small, and that glare varies in size and shape, a two-dimensional U-net structure is used as the model basis. An attention module is added to guide the model to focus on the glare region in the image. At the same time, a multi-scale module is added to improve the model's ability to capture glare at different scales, thereby improving the glare suppression effect of the intelligent model and achieving better high-quality CMS image imaging.

[0004] To achieve the above technical objectives, the present invention adopts the following technical solution: A method for suppressing glare in CMS images includes the following steps: S1: Establishment of learning samples for glare suppression in CMS images.

[0005] S2: Building and training an intelligent model for glare suppression in CMS images.

[0006] S3: Glare suppression for unknown CMS images.

[0007] Furthermore, the establishment of learning samples for glare suppression in CMS images in S1 specifically includes the following steps: S11, collect imaging images of the driving process and glare images mainly consisting of reflected glare and scattered glare.

[0008] S12 preprocesses the images of the driving scene, deleting images with low resolution or insufficient imaging information, and then selecting the glare-free images for subsequent label production.

[0009] S13: The glare-free driving scene image is randomly cropped to a size of (W, H). The glare image is cropped by opening a window of the same size (W, H) at the center of the glare. Then, the cropped glare image is randomly added to the cropped glare-free driving scene image to obtain the synthesized glare image.

[0010] S14, normalize the image with glare and the corresponding image without glare to [0,1]. This pair of images is a sample pair in the learning samples, with the former being the input feature and the latter being the label (ideal output).

[0011] It should be noted that during the process of randomly cropping images of glare-free driving scenes, images with insufficient imaging information may be obtained. Therefore, it is necessary to manually reduce the cropped images before synthesizing glare images.

[0012] Furthermore, the construction and training of the intelligent model for CMS image glare suppression in S2 specifically includes the following steps: S21. Construct a U-net using an encoder-decoder structure, with an input layer size of (W, H, 3), and determine the activation functions for each convolutional layer and output layer.

[0013] S22 adds a channel attention module after the last pooling layer of the encoder to capture the dependencies between channels, thereby weighting the differences in the original feature map.

[0014] S23, after the attention module, a multi-scale module is designed to achieve multi-scale feature extraction, and finally the obtained features of different scales are fused and passed to the decoder.

[0015] S24. Load the pre-trained deep convolutional neural network VGG16 model, select the feature extraction layer as the basis for calculating the perceptual loss, and then calculate the mean square error of the input image and the target image in the VGG feature space, which serves as the direction for network optimization.

[0016] Furthermore, glare suppression of the unknown CMS image in S3 specifically includes the following steps: S31, crop the unknown driving scene image to a size of (512, 512) and normalize it to [0, 1].

[0017] S32, input the above image into the trained network to obtain the anti-glare effect of the image.

[0018] Compared with the prior art, the technical advantages of the present invention are as follows: 1. Precisely focuses on the glare area, improving suppression efficiency.

[0019] To address the problem that glare pixels in CMS images have a small proportion and are easily ignored by the model in existing technologies, this invention introduces an attention module to guide the model to prioritize glare areas in the image, avoiding detection omissions due to low glare proportions and significantly improving the accuracy of glare recognition and suppression.

[0020] 2. It has strong multi-scale adaptability and covers different forms of glare.

[0021] To address the challenge of diverse glare sizes and shapes, this invention adds a multi-scale module, enabling the model to capture glare features at different scales (such as strong light spots, light patches, and scattered halos). This solves the problem of insufficient processing capability for complex glare shapes in traditional methods and improves the model's adaptability to changing driving scenarios.

[0022] 3. Optimize imaging quality and enhance scene robustness.

[0023] Based on the U-net architecture and the synergistic effect of the dual modules, this invention effectively overcomes glare interference caused by changes in light intensity and distribution during driving, achieving a more stable image deglare effect, thereby improving the imaging quality of CMS in changing driving environments and providing clearer external information for intelligent driving.

[0024] 4. The technical solution has a high degree of integration and is adapted to the needs of intelligent driving.

[0025] By combining modular design with deep learning models, this invention improves the intelligence level of CMS at the algorithm level without increasing hardware complexity, making it easier to integrate into intelligent driving systems and providing key technical support for electronic rearview mirrors to replace traditional glass rearview mirrors. Attached Figure Description

[0026] Figure 1 This is a flowchart of a CMS image glare suppression method provided in an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of the learning samples provided in the embodiments of the present invention, wherein parts (a) and (d) are the original images, which serve as the labels of the learning samples, and parts (b) and (e) are reflected glare and scattered glare, respectively. Parts (a) and (d) are superimposed with parts (b) and (e) respectively to obtain the glare images shown in parts (c) and (f), which serve as the original input of the model, i.e., image features.

[0028] Figure 3 This is a schematic diagram of the intelligent model of the U-net structure established for an embodiment of the present invention.

[0029] Figure 4 This is a schematic diagram of the learning loss curve provided for an embodiment of the present invention.

[0030] Figure 5 The images shown are illustrations of the effect of intelligent network glare suppression provided in the embodiments of the present invention. The left image is a glare image, the middle image is the original image without glare, and the right image is the glare reduction effect of the model established in the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0032] like Figure 1 As shown, the present invention proposes a CMS image glare suppression method, which includes the following steps: S1: Establishment of learning samples for glare suppression in CMS images.

[0033] S2: Building and training an intelligent model for glare suppression in CMS images.

[0034] S3: Glare suppression processing for unknown CMS images.

[0035] In practice, the establishment of learning samples for glare suppression in CMS images includes the following steps: S11, collect imaging images of the driving process and glare images mainly consisting of reflected glare and scattered glare.

[0036] S12 preprocesses the images of the driving scene, deleting images with low resolution or insufficient imaging information, and then selecting the glare-free images for subsequent label production.

[0037] S13, the glare-free driving scene image is randomly cropped to a size of (512, 512). The glare image is cropped by opening a window of the same size (512, 512) at the center of the glare. Then, the cropped glare-free driving scene image and the cropped glare image are randomly mixed to obtain a synthesized glare image.

[0038] S14, normalize the image with glare and the corresponding image without glare to [0,1]. This pair of images is a sample pair in the learning samples, with the former being the input feature and the latter being the label (ideal output).

[0039] It should be noted that during the process of randomly cropping images of glare-free driving scenes, images with insufficient imaging information may be obtained. Therefore, necessary manual reduction is required before synthesizing glare images.

[0040] like Figure 2 The figure shows the process of establishing learning samples according to an embodiment of the present invention. As can be seen from the figure, the image with glare is obtained by adding the corresponding non-glare image and the glare image in linear space.

[0041] In practical implementation, the construction and training of an intelligent model for CMS image glare suppression includes the following steps: S21. Construct a U-net using an encoder-decoder structure with an input size of (512, 512, 3), a ReLU activation function for each convolutional layer, and a sigmoid activation function for the output layer.

[0042] S22 adds a channel attention module after the last pooling layer of the encoder. It captures the dependencies between channels through global average pooling and global max pooling, and then generates attention weights through a fully connected layer, thereby weighting the differences in the original feature map.

[0043] S23. After the attention module, a multi-scale module is designed. This multi-scale module includes four different feature extraction methods: original size, 3×3 convolution, two 3×3 convolutions (equivalent to 5×5 convolution), and average pooling combined with 1×1 convolution. The features obtained at different scales will eventually be fused and passed to the decoder.

[0044] S24. Load the pre-trained VGG16 model, use the features of the block_conv3 layer as the basis for calculating the perceptual loss, and then calculate the mean square error of the input image and the target image in the VGG feature space, which serves as the direction for network optimization.

[0045] like Figure 3 The image shows a network with an encoder-decoder structure established in one embodiment of the present invention, which also includes an attention module and a multi-scale module.

[0046] like Figure 4 The diagram shows the loss curve (learning curve) during network training. It can be seen that as the number of iterations increases, the training loss and validation loss gradually decrease. However, when the number of iterations approaches 100, the loss value of the validation data fluctuates around 1.25, indicating that the model tends to fit.

[0047] In specific implementation, glare suppression processing for unknown CMS images includes the following steps: S31, crop the unknown driving scene image to a size of (512, 512) and normalize it to [0, 1].

[0048] S32, input the above image into the trained network to obtain the anti-glare effect of the image.

[0049] like Figure 5 To illustrate the glare suppression effect of the network constructed in one embodiment of the present invention on two unknown driving scene images, a comparison between the middle sub-image and the right image after glare suppression by the intelligent network shows that the driving scene glare suppression method proposed in this invention has good performance.

Claims

1. A method for suppressing glare in CMS images, characterized in that, Includes the following steps: S1: Establishment of learning samples for glare suppression in CMS images; S2: Building and training an intelligent model for glare suppression in CMS images; S3: Glare suppression for unknown CMS images.

2. The CMS image glare suppression method according to claim 1, characterized in that, The establishment of learning samples for glare suppression in CMS images in S1 specifically includes the following steps: S11, collect imaging images of the driving process and glare images, mainly reflected glare and scattered glare; S12, preprocess the images of the driving scene; S13, randomly crop the glare-free driving scene image with a size of (W, H), and crop the glare image by opening a window with the same size of (W, H) at the center of the glare. Then, randomly add the cropped glare image to the cropped glare-free driving scene image to obtain the synthesized glare image. S14, normalize the image with glare and the corresponding image without glare to [0,1]. This pair of images is a sample pair in the learning samples, with the former being the input feature and the latter being the label.

3. The CMS image glare suppression method according to claim 2, characterized in that, In S12, the preprocessing of images in driving scenes involves deleting images with low resolution or insufficient imaging information, and then selecting glare-free images from them for subsequent label creation.

4. The CMS image glare suppression method according to claim 1, characterized in that, The construction and training of the intelligent model for glare suppression in CMS images in S2 specifically includes the following steps: S21. Construct a U-net using an encoder-decoder structure with an input size of (W, H, 3), and determine the activation functions of its convolutional and output layers. S22, a channel attention module is added after the last pooling layer of the encoder to capture the dependencies between channels and to weight the differences in the original feature map. S23, after the attention module, a multi-scale module is designed to fuse the obtained features at different scales and pass them to the decoder; S24. Load the pre-trained VGG16 model, select the feature extraction layer as the basis for calculating the perceptual loss, and then calculate the mean square error of the input image and the target image in the VGG feature space, which serves as the direction for network optimization.

5. A CMS image glare suppression method according to claim 4, characterized in that, The size of the input layer of the U-net described in S21 is (W, H, 3).

6. The CMS image glare suppression method according to claim 1, characterized in that, Glare suppression of unknown CMS images in S3 specifically includes the following steps: S31, crop the unknown driving scene image and normalize it to [0,1]; S32, input the above image into the trained network to obtain the anti-glare effect of the image.

7. A CMS image glare suppression method according to claim 6, characterized in that, In S31, the image of the unknown driving scene is cropped to a size of (512, 512).

Images