Highway scene reconstruction method based on low-light image enhancement

By constructing a low-light image enhancement model of the lighting optimization and texture detail enhancement module, combining daytime and night image datasets, the image blurring and noise problems in night highway scenes are solved, and the accuracy of object detection and image segmentation effect are improved.

CN119006657AActive Publication Date: 2024-11-22ZHEJIANG UNIV OF TECH

Patent Information

Application Number
CN202410879566.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-02
Publication Date
2024-11-22
Estimated Expiration
2044-07-02

AI Technical Summary

Technical Problem

Image blurring and noise problems in night highway scenes lead to poor target detection results, difficult to distinguish background and target, and lack of paired low-light images-normal light image datasets for model training.

Method used

Using a method based on low light image enhancement, a low light image enhancement model including a light optimization module and a texture detail enhancement module is constructed, combined with the daytime and night highway image data sets for training, enhance image brightness and texture details, and form suitable shooting scenes. Changeable dataset.

Benefits of technology

It improves the visual quality of night images, adds effective image information, helps to more accurately identify and segment image boundaries, and improves the effect of object detection and segmentation tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119006657A_ABST
    Figure CN119006657A_ABST
Patent Text Reader

Abstract

The invention discloses an expressway scene reconstruction method based on low-light image enhancement, and the method comprises the steps: collecting a daytime expressway image and a real night expressway image, and synthesizing the daytime expressway image into a night expressway image; the synthesized night expressway image and the real night expressway image jointly serve as a data set, and the data set is divided into a training set and a test set; constructing a low-light image enhancement model, wherein the low-light image enhancement model comprises an illumination optimization module and a texture detail enhancement module; training the low-light image enhancement model by using the training set and testing by using the test set; and inputting an image to be reconstructed into the tested low-light image enhancement model, wherein the output final feature is an image reconstruction result. The method is beneficial for increasing effective information of the image and distinguishing the boundary of the image, and facilitates tasks such as image recognition, segmentation and detection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent transportation and computer vision, and specifically relates to a highway scene reconstruction method based on low-light image enhancement. Background Art

[0002] With the development of intelligent technology, the construction of smart highways has been very popular in recent years. The construction of smart highways is conducive to the high-quality development of highway traffic. However, there are also problems such as uneven lighting and blurred images in night environments. Therefore, in night environments, due to insufficient lighting on highways, road spills that drivers and passengers can find during the day are difficult to distinguish at night, which poses a great safety hazard and easily causes traffic accidents. The traffic accident rate at night is three times that of the day. In order to ensure the smooth progress of subsequent tasks in night highway scenes, a more robust detection algorithm is needed. At present, target detection in night highway scenes faces two major challenges. First, the difficulty of road detection tasks is further increased. Second, the night image enhancement task faces difficulties. For the second challenge, there are the following specific problems: (1) Due to the real-time changes in the shooting scene, it is impossible to obtain paired low-light image-normal light image datasets in real scenes for model training. (2) Night images contain a lot of noise and lose many features of different objects, resulting in poor results when used directly for detection, segmentation and other tasks, and it is difficult to distinguish between background and target. Summary of the invention

[0003] The purpose of the present invention is to address the above problems and propose a highway scene reconstruction method based on low-light image enhancement, which is beneficial to increase the effective information of the image and distinguish the image boundaries, and facilitate image recognition, segmentation, detection and other tasks.

[0004] To achieve the above object, the technical solution adopted by the present invention is:

[0005] The present invention proposes a highway scene reconstruction method based on low-light image enhancement, which comprises the following steps:

[0006] S1, collecting daytime highway images and real nighttime highway images, and synthesizing the daytime highway images into nighttime highway images;

[0007] S2, taking the synthetic nighttime highway images and the real nighttime highway images as a dataset, and dividing the dataset into a training set and a test set;

[0008] S3. Construct a low-light image enhancement model. The low-light image enhancement model includes an illumination optimization module and a texture detail enhancement module, wherein:

[0009] The illumination optimization module is used to iterate the input image and obtain the illumination optimization feature map I T, the learning function F(I t )The formula is as follows:

[0010]

[0011] Among them, u t is the illumination at the tth iteration, H(·) is the illumination function, I 0 is the input image of the 0th iteration, that is, the original input image S, I t is the input image of the tth iteration, I t+1 is the input image of the t+1th iteration, θ is the parameter of the illumination function, t∈{0, 1, …, T-1}, and T is the number of iterations;

[0012] The Texture Detail Enhancement module is used to perform the following operations:

[0013] Normal light characteristic diagram The Sobel operator is used to extract features in the height direction and width direction, and the corresponding height direction feature map is obtained. And width direction feature map Normal light characteristic diagram That is, the illumination optimization feature map I T ;

[0014] The height directional feature map And width direction feature map Superposition forms a first fusion feature map;

[0015] The first fusion feature map is combined with the normal light feature map Superimposing to form a second fusion feature map;

[0016] The second fusion feature map is convolved through the sixteenth convolution layer to obtain the final feature

[0017] S4, using the training set to train the low-light image enhancement model and using the test set to test it;

[0018] S5. Input the image to be reconstructed into the tested low-light image enhancement model, and the final feature output is This is the image reconstruction result.

[0019] Preferably, the illumination optimization module comprises a self-calibration module and an illumination construction module, wherein:

[0020] The constraint function G(I t ), the formula is as follows:

[0021]

[0022] Among them, R tis the reflection image of the tth iteration, is the dot division operation, s t is the difference between the input image of the tth iteration and the original input image, K is the image difference function, is the parameter of the image difference function, v t is the input image of the tth iteration after calibration, which is used as the input image of the tth iteration of the illumination construction module;

[0023] The illumination construction module includes a second addition operation, and a fourth feature extraction module, a fifth feature extraction module, a sixth feature extraction module, a seventh feature extraction module and an eighth feature extraction module connected in sequence, the fourth feature extraction module includes a sixth convolutional layer and a sixth activation function connected in sequence, the fifth feature extraction module includes a seventh convolutional layer, a fifth BN layer and a seventh activation function connected in sequence, the sixth feature extraction module includes an eighth convolutional layer, a sixth BN layer and an eighth activation function connected in sequence, the seventh feature extraction module includes a ninth convolutional layer, a seventh BN layer and a ninth activation function connected in sequence, the eighth feature extraction module includes a tenth convolutional layer and a tenth activation function connected in sequence, the input end of the sixth convolutional layer is also connected to the output end of the self-calibration module, and the second addition operation is used to superimpose the output features of the self-calibration module with the output features of the eighth feature extraction module as the input image of the t+1th iteration.

[0024] Preferably, the image difference function includes a first feature extraction module, a second feature extraction module and a third feature extraction module connected in sequence, the first feature extraction module includes a first convolutional layer, a first BN layer, a first activation function, a second convolutional layer, a second BN layer and a second activation function connected in sequence, the second feature extraction module includes a third convolutional layer, a third BN layer, a third activation function, a fourth convolutional layer, a fourth BN layer and a fourth activation function connected in sequence, and the third feature extraction module includes a fifth convolutional layer and a fifth activation function connected in sequence.

[0025] Preferably, the low-light image enhancement model also includes an image extraction module, which includes a ninth feature extraction module, a tenth feature extraction module, an eleventh feature extraction module, a twelfth feature extraction module and a thirteenth feature extraction module connected in sequence, the ninth feature extraction module includes an eleventh convolutional layer and an eleventh activation function connected in sequence, the tenth feature extraction module includes a twelfth convolutional layer, an eighth BN layer and a twelfth activation function connected in sequence, the eleventh feature extraction module includes a thirteenth convolutional layer, a ninth BN layer and a thirteenth activation function connected in sequence, the twelfth feature extraction module includes a fourteenth convolutional layer, a tenth BN layer and a fourteenth activation function connected in sequence, the thirteenth feature extraction module includes a fifteenth convolutional layer and a fifteenth activation function connected in sequence, the input feature of the eleventh convolutional layer is the original input image, and the output feature of the fifteenth convolutional layer is used as the input image of the first iteration.

[0026] Preferably, each activation function is a Relu activation function or a Sigmoid activation function.

[0027] Preferably, the Sobel operator performs the following operations:

[0028] Calculate normal light feature map The grayscale gradient value of each pixel point, then the grayscale gradient value Sobel(p) of the pth pixel point, is calculated as follows:

[0029]

[0030] Among them, Sobel h (p) is the gray gradient value in the height direction of the p-th pixel, Sobel w (p) is the grayscale gradient value of the p-th pixel in the width direction.

[0031] Determine whether the grayscale gradient value Sobel(p) of the p-th pixel is greater than the preset threshold. If so, the p-th pixel is considered to be an edge point and the corresponding image edge information is obtained. Otherwise, the p-th pixel is considered not to be an edge point and there is no need to obtain the corresponding image edge information.

[0032] Preferably, the convolution kernel size of the Sobel operator is 3×3.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] This method can adapt to the real-time changes of the shooting scene by mixing the data set to form a paired low-light image-normal light image data set, which is convenient for the training of the low-light image enhancement model, especially by constructing a low-light image enhancement model including an illumination optimization module and a texture detail enhancement module. The illumination optimization module can enhance the brightness of night images, enrich image features, and increase image effective information. The texture detail enhancement module can increase image texture detail features, making the information at the edge of the image richer, which is conducive to extracting the edges of objects in the image and avoiding the problem that night images contain a lot of noise and lose the features of many different objects. It is helpful for subsequent tasks to distinguish image boundaries, locate targets more accurately, and distinguish target edge features from background, so as to facilitate image recognition, segmentation, detection and other tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flow chart of the highway scene reconstruction method based on low-light image enhancement of the present invention;

[0036] Figure 2 It is a structural schematic diagram of the low-light-level image enhancement model of the present invention;

[0037] Figure 3 It is a structural schematic diagram of the illumination optimization module and the image extraction module of the present invention;

[0038] Figure 4 It is a structural schematic diagram of the texture detail enhancement module of the present invention. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0040] It should be noted that when a module is referred to as being "connected" to another module, it may be directly connected to the other module or there may be a central module. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present application. The terms used herein in the specification of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application.

[0041] like Figure 1-4 As shown, a highway scene reconstruction method based on low-light image enhancement includes the following steps:

[0042] S1. Collect daytime highway images and real nighttime highway images, and synthesize the daytime highway images into nighttime highway images.

[0043] S2. The synthetic nighttime highway images and the real nighttime highway images are taken as a dataset, and the dataset is divided into a training set and a test set.

[0044] Among them, by collecting daytime highway images, Adobe Lightroom software can be used to artificially synthesize nighttime highway images by changing the exposure, highlight, shadow, saturation, and white level attributes of the daytime highway images. Together with the collected real nighttime highway images, they are used as a data set for the subsequent training and testing of the low-light image enhancement model.

[0045] S3. Construct a low-light image enhancement model. The low-light image enhancement model includes an illumination optimization module and a texture detail enhancement module, wherein:

[0046] The illumination optimization module is used to iterate the input image and obtain the illumination optimization feature map I T , the learning function F(I t )The formula is as follows:

[0047]

[0048] Among them, u t is the illumination at the tth iteration, H(·) is the illumination function, I 0 is the input image of the 0th iteration, that is, the original input image S, I t is the input image of the tth iteration, I t+1 is the input image of the t+1th iteration, θ is the parameter of the illumination function, t∈{0, 1, …, T-1}, and T is the number of iterations;

[0049] The Texture Detail Enhancement module is used to perform the following operations:

[0050] Normal light characteristic diagram The Sobel operator is used to extract features in the height direction and width direction, and the corresponding height direction feature map is obtained. And width direction feature map Normal light characteristic diagram That is, the illumination optimization feature map I T ;

[0051] The height directional feature map And width direction feature map Superposition forms a first fusion feature map;

[0052] The first fusion feature map is combined with the normal light feature map Superimposing to form a second fusion feature map;

[0053] The second fusion feature map is convolved through the sixteenth convolution layer to obtain the final feature

[0054] Among them, Figure 3 As shown in FIG. 1 , the illumination optimization module extracts features from the original input image (which may be a synthetic nighttime highway image or a real nighttime highway image) to obtain the output features of the current iteration, and then inputs the output features of the current iteration into the illumination optimization module for the next iteration. After T iterations, the illumination optimization feature map I is obtained. T .

[0055] F(I t ) is the learning function. The whole learning process is iterated T times, t∈{0, 1, …, T-1}, θ is the parameter of the illumination function, which is specifically expressed as the weight in the convolution process of the illumination function.

[0056] like Figure 2 As shown, the input feature of the texture detail enhancement module (the internal structure is not fully shown in the figure) is the normal light feature map (i.e., illumination optimization feature map I T ),like Figure 4 As shown, from the normal light characteristic diagram The Sobel operator is used to extract edge texture details in the width and height directions to obtain the height direction feature map And width direction feature map Furthermore, the height directional feature map And width direction feature map The first fusion feature map is superimposed, and then the first fusion feature map is combined with the normal feature map After superposition, the final features are obtained by further fusion through convolution operation.

[0057] Specifically, the final feature The acquisition is as follows:

[0058]

[0059] in, is the feature map obtained after the illumination optimization module, Conv 3×3 It is a convolution layer with a convolution kernel size of 3×3, that is, the sixteenth convolution layer.

[0060] In one embodiment, the Sobel operator performs the following operations:

[0061] Calculate normal light feature map The grayscale gradient value of each pixel point, then the grayscale gradient value Sobel(p) of the pth pixel point, is calculated as follows:

[0062]

[0063] Among them, Sobel h (p) is the gray gradient value in the height direction of the p-th pixel, Sobel w (p) is the grayscale gradient value of the p-th pixel in the width direction.

[0064] Determine whether the grayscale gradient value Sobel(p) of the p-th pixel is greater than the preset threshold. If so, the p-th pixel is considered to be an edge point and the corresponding image edge information is obtained. Otherwise, the p-th pixel is considered not to be an edge point and there is no need to obtain the corresponding image edge information.

[0065] In one embodiment, the convolution kernel size of the Sobel operator is 3×3.

[0066] The Sobel operator uses a 3×3 convolution kernel to calculate the grayscale gradient value of each pixel in the height direction and width direction of the image, and performs square sum operation on the grayscale gradient values ​​in the two directions to obtain the grayscale gradient value of the pixel. If the grayscale gradient value of the corresponding pixel is greater than the preset threshold, the point is considered to be an edge point, thereby obtaining the corresponding image edge information. Otherwise, there is no need to obtain it.

[0067] By adding the Sobel operator to process the reconstructed image features (normal light feature map ), the details of the reconstructed image features are enhanced to obtain an enhanced image, which can improve the visual quality of the image and help subsequent detection and classification tasks to more accurately locate the target and distinguish the target edge features and background.

[0068] S4. Use the training set to train the low-light image enhancement model and use the test set to test it.

[0069] The training samples in the training set are input into the low-light image enhancement model for training, the loss is calculated, and the network parameters are updated. After the training is completed, the test is performed to obtain a tested low-light image enhancement model.

[0070] S5. Input the image to be reconstructed into the tested low-light image enhancement model, and the final feature output is This is the image reconstruction result.

[0071] After training and testing are completed, the image to be reconstructed is input into the tested low-light image enhancement model for processing, and the image reconstruction result, i.e. the final feature

[0072] This method can adapt to the real-time changes of the shooting scene by mixing the data set to form a paired low-light image-normal light image data set, which is convenient for model training. The low-light image enhancement model includes an illumination optimization module and a texture detail enhancement module. The illumination optimization module can enhance the brightness of night images, enrich image features, and increase the effective information of the image. The texture detail enhancement module can increase the image texture detail features, making the information at the edge of the image richer, which is conducive to extracting the edges of objects in the image, avoiding the problem that night images contain a lot of noise and lose the features of many different objects, and helping to distinguish the image boundaries in subsequent tasks, which is more conducive to image recognition, segmentation, detection and other tasks.

[0073] In one embodiment, the illumination optimization module includes a self-calibration module and an illumination construction module, wherein:

[0074] The constraint function G(I t ), the formula is as follows:

[0075]

[0076] Among them, R t is the reflection image of the tth iteration, is the dot division operation, s t is the difference between the input image of the tth iteration and the original input image, K is the image difference function, is the parameter of the image difference function, v t is the input image of the tth iteration after calibration, which is used as the input image of the tth iteration of the illumination construction module;

[0077] The illumination construction module includes a second addition operation, and a fourth feature extraction module, a fifth feature extraction module, a sixth feature extraction module, a seventh feature extraction module and an eighth feature extraction module connected in sequence, the fourth feature extraction module includes a sixth convolutional layer and a sixth activation function connected in sequence, the fifth feature extraction module includes a seventh convolutional layer, a fifth BN layer and a seventh activation function connected in sequence, the sixth feature extraction module includes an eighth convolutional layer, a sixth BN layer and an eighth activation function connected in sequence, the seventh feature extraction module includes a ninth convolutional layer, a seventh BN layer and a ninth activation function connected in sequence, the eighth feature extraction module includes a tenth convolutional layer and a tenth activation function connected in sequence, the input end of the sixth convolutional layer is also connected to the output end of the self-calibration module, and the second addition operation is used to superimpose the output features of the self-calibration module with the output features of the eighth feature extraction module as the input image of the t+1th iteration.

[0078] Among them, the illumination calibration feature map (i.e., the input image v of the tth iteration after calibration) is obtained through the self-calibration module t), and then input the illumination calibration feature map into the illumination construction module to obtain the illumination construction feature map (i.e., the illumination optimization feature map of the tth iteration). The self-calibration module is used to calculate the difference between the illumination construction feature map and the original input image in each iteration, and use these differences to constrain the model training process. K is the image difference function, which represents the convolution process of the self-calibration module. is the parameter of the image difference function, that is, the weight in the convolution process.

[0079] The use of the self-calibration module and the illumination construction module can enhance the brightness of night highway images (which can be synthetic night highway images or real night highway images), enrich image features, and increase image effective information, which is beneficial to other subsequent image tasks, such as detection, segmentation and other tasks.

[0080] In one embodiment, the image difference function includes a first feature extraction module, a second feature extraction module and a third feature extraction module connected in sequence, the first feature extraction module includes a first convolutional layer, a first BN layer, a first activation function, a second convolutional layer, a second BN layer and a second activation function connected in sequence, the second feature extraction module includes a third convolutional layer, a third BN layer, a third activation function, a fourth convolutional layer, a fourth BN layer and a fourth activation function connected in sequence, and the third feature extraction module includes a fifth convolutional layer and a fifth activation function connected in sequence.

[0081] In one embodiment, the low-light image enhancement model also includes an image extraction module, which includes a ninth feature extraction module, a tenth feature extraction module, an eleventh feature extraction module, a twelfth feature extraction module and a thirteenth feature extraction module connected in sequence, the ninth feature extraction module includes an eleventh convolutional layer and an eleventh activation function connected in sequence, the tenth feature extraction module includes a twelfth convolutional layer, an eighth BN layer and a twelfth activation function connected in sequence, the eleventh feature extraction module includes a thirteenth convolutional layer, a ninth BN layer and a thirteenth activation function connected in sequence, the twelfth feature extraction module includes a fourteenth convolutional layer, a tenth BN layer and a fourteenth activation function connected in sequence, the thirteenth feature extraction module includes a fifteenth convolutional layer and a fifteenth activation function connected in sequence, the input feature of the eleventh convolutional layer is the original input image, and the output feature of the fifteenth convolutional layer is used as the input image of the first iteration.

[0082] In one embodiment, each activation function is a Relu activation function or a Sigmoid activation function.

[0083] in, Figure 3Each box in the figure represents the corresponding feature map obtained, and the arrow represents the operation process (i.e., the corresponding feature extraction module). The image difference function of the self-calibration module includes 5 convolutional layers, 4 BN layers, 4 ReLu layers (first to fourth activation functions) and 1 sigmoid function (fifth activation function), and the illumination construction module includes 5 convolutional layers, 3 BN layers, 4 ReLU layers (sixth to ninth activation functions) and 1 sigmoid function (tenth activation function). The image extraction module includes 5 convolutional layers, 3 BN layers, 4 ReLu layers (eleventh to fourteenth activation functions) and 1 sigmoid function (fifteenth activation function).

[0084] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0085] The above-described embodiments only express the more specific and detailed embodiments described in this application, but they cannot be understood as limiting the scope of the patent application. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of this application, which all belong to the protection scope of this application. Therefore, the protection scope of the patent application shall be based on the attached claims.

Claims

1. A highway scene reconstruction method based on low-light image enhancement, characterized by: The highway scene reconstruction method based on low-light image enhancement comprises the following steps: S1, collecting daytime highway images and real nighttime highway images, and synthesizing the daytime highway images into nighttime highway images; S2, taking the synthetic nighttime highway images and the real nighttime highway images as a dataset, and dividing the dataset into a training set and a test set; S3. Construct a low-light image enhancement model, wherein the low-light image enhancement model includes a lighting optimization module and a texture detail enhancement module, wherein: The illumination optimization module is used to iterate the input image in a loop to obtain an illumination optimization feature map I T , the learning function F(I t )The formula is as follows: Among them, u t is the illumination at the tth iteration, H(·) is the illumination function, I 0 is the input image of the 0th iteration, that is, the original input image S, I t is the input image of the tth iteration, I t+1 is the input image of the t+1th iteration, θ is the parameter of the illumination function, t∈{0, 1, …, T-1}, and T is the number of iterations; The texture detail enhancement module is used to perform the following operations: Normal light characteristic diagram The Sobel operator is used to extract features in the height direction and width direction, and the corresponding height direction feature map is obtained. And width direction feature map The normal light characteristic diagram That is, the illumination optimization feature map I T ; The height directional feature map And width direction feature map Superposition forms a first fusion feature map; The first fusion feature map is combined with the normal light feature map Superimposing to form a second fusion feature map; The second fusion feature map is convolved through the sixteenth convolution layer to obtain the final feature S4, using the training set to train the low-light image enhancement model and using the test set to test it; S5. Input the image to be reconstructed into the tested low-light image enhancement model, and the final feature output is This is the image reconstruction result.

2. The highway scene reconstruction method based on low-light image enhancement according to claim 1, characterized in that: The illumination optimization module includes a self-calibration module and an illumination construction module, wherein: The constraint function G(I t ), the formula is as follows: Among them, R t is the reflection image of the tth iteration, is the point division operation, s t is the difference between the input image of the tth iteration and the original input image, K is the image difference function, is the parameter of the image difference function, v t is the input image of the tth iteration after calibration, which is used as the input image of the tth iteration of the illumination construction module; The illumination construction module includes a second addition operation, and a fourth feature extraction module, a fifth feature extraction module, a sixth feature extraction module, a seventh feature extraction module and an eighth feature extraction module connected in sequence, the fourth feature extraction module includes a sixth convolutional layer and a sixth activation function connected in sequence, the fifth feature extraction module includes a seventh convolutional layer, a fifth BN layer and a seventh activation function connected in sequence, the sixth feature extraction module includes an eighth convolutional layer, a sixth BN layer and an eighth activation function connected in sequence, the seventh feature extraction module includes a ninth convolutional layer, a seventh BN layer and a ninth activation function connected in sequence, the eighth feature extraction module includes a tenth convolutional layer and a tenth activation function connected in sequence, the input end of the sixth convolutional layer is also connected to the output end of the self-calibration module, and the second addition operation is used to superimpose the output features of the self-calibration module with the output features of the eighth feature extraction module as the input image of the t+1th iteration.

3. The highway scene reconstruction method based on low-light image enhancement according to claim 2, characterized in that: The image difference function includes a first feature extraction module, a second feature extraction module and a third feature extraction module connected in sequence, the first feature extraction module includes a first convolutional layer, a first BN layer, a first activation function, a second convolutional layer, a second BN layer and a second activation function connected in sequence, the second feature extraction module includes a third convolutional layer, a third BN layer, a third activation function, a fourth convolutional layer, a fourth BN layer and a fourth activation function connected in sequence, and the third feature extraction module includes a fifth convolutional layer and a fifth activation function connected in sequence.

4. The highway scene reconstruction method based on low-light image enhancement according to claim 1, characterized in that: The low-light image enhancement model also includes an image extraction module, which includes a ninth feature extraction module, a tenth feature extraction module, an eleventh feature extraction module, a twelfth feature extraction module and a thirteenth feature extraction module connected in sequence, the ninth feature extraction module includes an eleventh convolutional layer and an eleventh activation function connected in sequence, the tenth feature extraction module includes a twelfth convolutional layer, an eighth BN layer and a twelfth activation function connected in sequence, the eleventh feature extraction module includes a thirteenth convolutional layer, a ninth BN layer and a thirteenth activation function connected in sequence, the twelfth feature extraction module includes a fourteenth convolutional layer, a tenth BN layer and a fourteenth activation function connected in sequence, the thirteenth feature extraction module includes a fifteenth convolutional layer and a fifteenth activation function connected in sequence, the input feature of the eleventh convolutional layer is the original input image, and the output feature of the fifteenth convolutional layer is used as the input image of the first iteration.

5. The highway scene reconstruction method based on low-light image enhancement according to any one of claims 2 to 4, characterized in that: Each of the activation functions is a Relu activation function or a Sigmoid activation function.

6. The highway scene reconstruction method based on low-light image enhancement according to claim 1, characterized in that: The Sobel operator performs the following operations: Calculate normal light feature map The grayscale gradient value of each pixel point, then the grayscale gradient value Sobel(p) of the pth pixel point, is calculated as follows: Among them, Sobel h (p) is the gray gradient value in the height direction of the p-th pixel, Sobel w (p) is the grayscale gradient value of the p-th pixel in the width direction. Determine whether the grayscale gradient value Sobel(p) of the p-th pixel is greater than the preset threshold. If so, the p-th pixel is considered to be an edge point and the corresponding image edge information is obtained. Otherwise, the p-th pixel is considered not to be an edge point and there is no need to obtain the corresponding image edge information.

7. The highway scene reconstruction method based on low-light image enhancement according to claim 6, characterized in that: The convolution kernel size of the Sobel operator is 3×3.

Citation Information

Patent Citations

  • Fast low-light image enhancement method based on global frequency domain filtering

    CN116777776A

  • Low-light scene image enhancement method and system based on scene self-adaption and self-distillation

    CN116894934A

  • Generating texture mesh using one or more neural networks

    CN117635871A

  • Automatic vehicle equipment monitoring, warning, and control system

    US20120050074A1

  • System and method for managing communication and / or storage of image data

    WO2006052390A2

Cited By

  • Object recognition method and related device

    CN121074530A

  • Object recognition method and related apparatus

    CN121074530B