A tunnel entrance image enhancement method based on group attention kernel feedback perception

Abstract

This invention discloses a tunnel entrance image enhancement method based on group attention kernel feedback perception. It includes an encoding module, constructing a kernel feedback perception network, constructing a decoding network, constructing a group attention feedback perception tunnel entrance image enhancement network, constructing a cost function, obtaining the optimal parameters of the deep network, and obtaining the enhanced tunnel entrance image based on the optimal network parameters. This invention can effectively improve the visual quality of tunnel entrance images, reduce noise in the acquired images, is simple and effective, has a short processing time, and can achieve real-time processing.

Description

Technical Field

[0001] This invention is applied to the background of computer vision and is entitled "A Method for Enhancing Tunnel Entrance Images Based on Group Attention Kernel Feedback Perception". Background Technology

[0002] With the rapid development of intelligent transportation, the application of visual methods has expanded. However, in areas with poor lighting conditions, such as tunnel entrances, visual acquisition is prone to degradation phenomena, such as low contrast, color cast, and noise. These degradations severely affect the extraction of important information from the acquired images, thus hindering further image understanding. Currently, processing these degraded images has become a significant challenge in the field of computer vision, especially in image enhancement. Image enhancement either sharpens unclear areas or highlights certain features of interest while suppressing uninteresting features. Both strategies can improve image quality and provide reliable visual features for subsequent image understanding tasks such as target tracking and image segmentation.

[0003] To improve image quality in poorly lit conditions, researchers have made many successful attempts and achieved satisfactory results. Huang et al. used Gamma correction to increase the brightness of dark areas and decrease the brightness of brighter areas to improve the low contrast of low-light images. The Retinex image model proposed by Land has also gradually become an effective method for image enhancement. This method improves image quality by decomposing the illumination layer and the reflection layer, but how to effectively decompose them is currently the key to solving this type of problem. One approach is to apply certain prior constraints to the illumination layer or the reflection layer, or both. Common methods include bilateral filtering, L2 and L1 penalty terms on the gradient, and total variation (TV) minimization. Another approach is to use deep learning to separate the illumination layer or the reflection layer. Shen et al. improved the accuracy of decomposition and image quality by designing a deep learning network that combines the traditional multi-scale Retinex method. Gharbi et al. used an enhanced bilateral learning network to predict the pointwise color transformation coefficients of the color vector at each pixel. While these algorithms effectively solve the separation problem of the illumination layer or reflection layer end-to-end from deep learning methods and avoid imposing certain prior constraints on the illumination layer or reflection layer, they ignore the influence of noise when processing these images. If not handled properly, they may make incorrect estimates of the illumination layer and reflection layer in darker areas, affecting image quality. To improve the image quality of tunnels under noisy, low-light conditions, this invention introduces visual attention into deep networks and proposes a novel tunnel entrance image enhancement network based on group attention and kernel feedback perception. This network improves the image quality of tunnel entrances through group attention and kernel feedback perception methods, solving the problem of low-light, noisy tunnel entrance images, improving image contrast, and enhancing image details. Summary of the Invention

[0004] In complex backgrounds, especially under low-light conditions, acquired images often suffer from noise and low-contrast degradation. Common image enhancement methods, while reducing noise, often blur the image, resulting in the loss of some details. Other methods, while improving contrast in low-contrast situations, introduce new noise during processing. To address this issue, this invention provides a tunnel entrance image enhancement method based on group attention kernel feedback perception. This method effectively improves the quality of tunnel entrance images in complex backgrounds, reduces image noise, enhances image contrast, and provides robust feature information for further image understanding. The method is simple, effective, and has a short processing time, enabling real-time image enhancement.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a tunnel entrance image enhancement method based on group attention kernel feedback perception, comprising the following steps:

[0006] Construct an encoding module; construct a kernel feedback perception network; construct a decoding network; construct a tunnel entrance image enhancement network with group attention kernel feedback; construct a cost function, obtain the optimal parameters of the deep network, and obtain the enhanced tunnel entrance image based on the optimal network parameters.

[0007] In one embodiment, the construction encoding module is composed of two cascaded group attention convolutional modules. Each group attention convolutional module is composed of two convolutional blocks and one cascaded group attention module; wherein, each convolutional block is composed of a convolutional layer with a kernel width of 3, a batch processing layer and a ReLU activation function; wherein, the mathematical form of the convolutional block is composed of equation (1).

[0008] (1)

[0009] Where F CB This represents the visual features after the input visual features X have passed through the convolutional block. The convolutional layer function with a kernel width of 3 is represented by BN(.), the batch layer function is represented by BN(.), and the modified linear unit is represented by ReLU(.); the mathematical form of each group attention module is composed of equation (2).

[0010] (2)

[0011] in Yes The i-th visual feature is segmented along visual feature channel C. h, w, and C represent the length, width, and number of channels of visual feature X, respectively. c1 represents the number of channels in each segmented block, with a size of C / n, where n is the number of blocks to be segmented. DCT(.) represents the Discrete Cosine Transform function, and Concat(.) represents each segmented visual feature X. i After discrete cosine transform, the concatenation operation of features is performed, where FC(.) represents the fully connected layer function and ReLU(.) represents the modified linear unit. This represents the activation function, specifically the sigmoid function in this case. This indicates channel-level multiplication, with larger coefficients assigned to important channels, F. MGCA It is the visual feature X after passing through the group attention module. The group attention convolution module combines the group attention information and the convolution block information to form a new visual representation module with stronger capabilities. Its mathematical expression is specifically composed of equation (3).

[0012] (3)

[0013] Where X is the input feature X, ConvB(.) and MGCA(.) represent the functions of the visual features after passing through the convolutional block and the group attention module, respectively, and F GAC This represents the visual features after passing through the group attention convolution module.

[0014] In one embodiment, the construction of the kernel feedback perception module involves combining the adaptive adjustment of the fuzzy kernel of the convolutional layer with the kernel feedback perception information of the next layer to effectively improve the visual feature representation capability. The kernel adaptive correction network mainly includes two branches: the kernel adaptive correction network and the group attention convolutional deep feature extraction network. The initial kernel estimation strategy in the kernel adaptive correction network adopts a method similar to the kernel stretching method used by Kaizhang in the super-resolution algorithm SRMD. The kernel adaptive correction network first obtains a noisy feature map through a dimension stretching method, followed by two convolutional blocks and one 3x3 convolutional layer. The group attention convolutional deep feature extraction network first passes through the group attention convolutional module, followed by a deep feature extraction block. After the two parallel branches are completed, the visual features after the two parallel branches are concatenated together, and then the concatenated visual features are reduced to the same channel dimension as the initial features through a 1x1 convolutional layer. Subsequently, through short links, the visual features of the previous layer and the visual features after a series of module operations are added pixel-wise to obtain the input features of the next layer kernel feedback perception network. The mathematical form of the kernel feedback sensing network is constituted by equation (4).

[0015]

[0016]

[0017] (4)

[0018] Where Conv 3x3 This represents a convolutional layer with a kernel width of 3. 1x1 A convolutional layer with a kernel width of 1 is represented by ConvB2(.), which represents the convolutional module function. The subscript 2 indicates that the convolutional module function is executed twice. DS(.) represents the dimension stretching module function, and C(.) represents the cascading operation. This represents element-wise addition. and These represent the output features after passing through the two branches. DFB(.) represents the deep feature extraction function, which is mainly composed of convolutional layers and pooling layers. The input features of this module are passed through a convolutional layer with a kernel width of 1 and a pooling layer with a kernel width of 3, and then concatenated. After the same operation, they are concatenated again, and then linearly fused through a convolutional layer with a kernel width of 1 to form a new visual feature output, which is composed of Equation (5).

[0019]

[0020]

[0021] (5)

[0022] in, These represent the visual features after passing through the first convolutional layer with a kernel width of 1, the first pooling layer with a kernel width of 3, and the second pooling layer, respectively. C represents the cascade operation. Conv 1x1 Pool represents a convolutional layer with a kernel width of 1. 3x3 X represents a pooling layer with a kernel width of 3. in and X out These represent the input and output characteristics of the module, respectively.

[0023] In one embodiment, the construction of the decoding network consists of two convolutional layers with edge supplementation, specifically expressed by equation (6).

[0024] (6)

[0025] Among them, F d It is the visual feature after two convolutions with edge padding. Here, X is the output feature after m kernel feedbacks. pConv2(.) represents the convolutional layer with edge padding, and the subscript 2 indicates that it is repeated twice.

[0026] In one embodiment, the construction group pays attention to the kernel feedback perception tunnel entrance image enhancement network. The network is mainly composed of a feature encoding module, a kernel feedback module, and a feature decoding module. The enhancement network first extracts visual features through the feature encoding module, then performs m kernel self-feedback feature processing, and then performs feature decoding to obtain high-frequency information of visual features. Finally, through short links, the contour information of the input features is combined with the high-frequency information to form a tunnel entrance image with satisfactory visual quality, providing reliable visual information for further image understanding. Its specific description is composed of Equation (7).

[0027] (7)

[0028] Where F_ out This represents the enhanced tunnel entrance image. `Encode(.)` and `Decode(,)` represent the feature encoding and decoding functions, respectively. KFAB m (.) indicates that the kernel feedback function has been applied m times. This indicates pixel-level addition.

[0029] In one embodiment, the deep network training uses a widely adopted adaptive momentum gradient descent algorithm network optimizer. During training, the learning rate is set to 0.0005~0.005, the momentum is set to 0.9, and the cost function used is the L1 cost function of the first paradigm. The network is effectively trained on internationally recognized datasets and a self-constructed tunnel entrance image dataset to obtain the optimal parameters of the augmented network.

[0030] Compared with existing technologies, the beneficial effects achieved by this invention are its simplicity and effectiveness, short processing time, and real-time processing capability. It is computationally simple, efficient, and highly feasible, making it easily applicable to personal computers and embedded systems. This invention can be applied to several fields of computer vision, such as object detection, object tracking, and object recognition. Attached Figure Description

[0031] The following detailed description of the specific embodiments of the present application in conjunction with the accompanying drawings will make the technical solutions and other beneficial effects of the present application apparent.

[0032] In the attached diagram:

[0033] Figure 1 This is a flowchart of the enhanced system structure for implementing the present invention;

[0034] Figure 2 This is a schematic diagram of the implementation process of the convolutional block module provided in an embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram of the implementation process of the group attention perception module provided in an embodiment of the present invention;

[0036] Figure 4 This is a schematic diagram of the implementation process of the group attention convolution module provided in an embodiment of the present invention;

[0037] Figure 5 This is a schematic diagram of the group attention kernel feedback module structure provided in an embodiment of the present invention;

[0038] Figure 6 This is a schematic diagram of the structure and flow of the deep feature extraction module provided in an embodiment of the present invention;

[0039] Figure 7 This is a schematic diagram of the group attention kernel feedback tunnel entrance image enhancement network structure provided in an embodiment of the present invention; Detailed Implementation

[0040] The following disclosure provides many different embodiments or examples for implementing different structures of this application. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the scope of this application. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. In addition, various specific examples of processes and materials are provided in this application, but those skilled in the art will recognize the application of other processes and / or the use of other materials.

[0041] Please see Figure 1-7 The present invention provides a technical solution: a method for enhancing tunnel entrance images based on group attention kernel feedback perception, the method comprising:

[0042] Step S01: Construct the training image set and validation image set required for the image augmentation network;

[0043] Step S02: Construct a group attention kernel feedback perception tunnel entrance image enhancement network;

[0044] Step S03: Combine the training image set and cost function to obtain the optimal parameters of the augmentation network and give the augmented tunnel entrance image.

[0045] Construct an encoding module; construct a kernel feedback perception network; construct a decoding network; construct a tunnel entrance image enhancement network with group attention kernel feedback; construct a cost function, obtain the optimal parameters of the deep network, and obtain the enhanced tunnel entrance image based on the optimal network parameters.

[0046] In one embodiment, the construction encoding module is composed of two cascaded group attention convolutional modules. Figure 2It can be observed that each group attention convolutional module consists of two convolutional blocks and one group attention module cascaded together; each convolutional block comprises a convolutional layer with a kernel width of 3, a batch processing layer, and a ReLU activation function, as shown below. Figure 3 As shown, the mathematical representation of a convolutional block is as follows:

[0047]

[0048] Where F CB This represents the visual features after the input visual features X have passed through the convolutional block. This represents a convolutional layer function with a kernel width of 3, BN(.) represents a batch layer function, and ReLU(.) represents a corrected linear unit. The structure of each group's attention module is as follows: Figure 4 As shown, its specific mathematical representation is as follows:

[0049]

[0050] in Yes The i-th visual feature is segmented along visual feature channel C. h, w, and C represent the length, width, and number of channels of visual feature X, respectively. c1 represents the number of channels in each segmented block, with a size of C / n, where n is the number of blocks to be segmented. DCT(.) represents the Discrete Cosine Transform function, and Concat(.) represents each segmented visual feature X. i After discrete cosine transform, the concatenation operation of features is performed, where FC(.) represents the fully connected layer function and ReLU(.) represents the modified linear unit. This represents the activation function, specifically the sigmoid function in this case. This indicates channel-level multiplication, with larger coefficients assigned to important channels, F. MGCA This refers to the visual feature X after passing through the group attention module. The group attention convolution module combines group attention information and convolutional block information to form a new, more powerful visual representation module, which is mathematically expressed as follows:

[0051]

[0052] Where X is the input feature X, ConvB(.) and MGCA(.) represent the functions of the visual features after passing through the convolutional block and the group attention module, respectively, and F GAC This represents the visual features after passing through the group attention convolution module.

[0053] In one embodiment, the constructed kernel feedback perception module combines the continuous adaptive adjustment of the fuzzy kernel of the convolutional layer with the group attention convolutional depth features to effectively improve the visual feature representation capability. Its network structure is as follows: Figure 5 As shown, it mainly consists of two branches: a kernel adaptive correction network and a group attention convolutional deep feature extraction network. The initial kernel estimation strategy in the kernel adaptive correction network adopts a method similar to kernel stretching used by Kaizhang in the super-resolution algorithm SRMD. The kernel adaptive correction network first obtains a noisy feature map through a dimensionality stretching method, followed by two convolutional blocks and one 3x3 convolutional layer. The group attention convolutional deep feature extraction network first passes through a group attention convolutional module, followed by a deep feature extraction block. After the two parallel branches, the visual features obtained from the two parallel branches are concatenated together, and then a 1x1 convolutional layer is used to reduce the concatenated visual features to visual features with the same channel dimension as the initial features. Subsequently, through short links, the visual features of the previous layer and the visual features after a series of module operations are added pixel-wise to obtain the input features of the next layer, the kernel feedback perceptual network. The mathematical form of the kernel feedback perceptual network is specifically represented as follows:

[0054]

[0055]

[0056]

[0057] Where Conv 3x3 This represents a convolutional layer with a kernel width of 3. 1x1 A convolutional layer with a kernel width of 1 is represented by ConvB2(.), which represents the convolutional module function. The subscript 2 indicates that the convolutional module function is executed twice. DS(.) represents the dimension stretching module function, and C(.) represents the cascading operation. This represents element-wise addition. and These represent the output features after the two branches. DFB(.) represents the deep feature extraction function, and its structure is as follows: Figure 6 As shown in the diagram, the input features of this module are passed through a convolutional layer with a kernel width of 1 and a pooling layer with a kernel width of 3, then concatenated. This process is repeated, followed by another concatenation, and finally linearly fused through a convolutional layer with a kernel width of 1 to form a new visual feature output. Specifically, this is represented as follows:

[0058]

[0059]

[0060]

[0061]

[0062] in, These represent visual features after passing through the first convolutional layer with a kernel width of 1, the first pooling layer with a kernel width of 3, and the second pooling layer, respectively. C represents the cascade operation. Conv 1x1 Pool represents a convolutional layer with a kernel width of 1. 3x3 X represents a pooling layer with a kernel width of 3. in and X out These represent the input and output characteristics of the module, respectively.

[0063] In one embodiment, the decoding network is constructed from two convolutional layers with edge padding, as described below:

[0064]

[0065] Among them, F d It is the visual feature after two convolutions with edge padding. Here, X is the output feature after m kernel feedbacks. pConv2(.) represents the convolutional layer with edge padding, and the subscript 2 indicates that it is repeated twice.

[0066] In one embodiment, the construction group attention kernel feedback-aware tunnel entrance image enhancement network is constructed. This network mainly consists of a feature encoding module, a kernel feedback module, and a feature decoding module, as follows: Figure 7 As shown, this augmentation network first extracts visual features through a feature encoding module, then performs m rounds of kernel self-feedback feature processing, followed by feature decoding to obtain high-frequency information of the visual features. Finally, through short links, the contour information of the input features is combined with the high-frequency information to form a tunnel entrance image with satisfactory visual quality, providing reliable visual information for further image understanding. Its mathematical form is specifically expressed as follows:

[0067]

[0068] Where F_ out This represents the enhanced tunnel entrance image. `Encode(.)` and `Decode(,)` represent the feature encoding and decoding functions, respectively. KFAB m (.) indicates that the kernel feedback function has been applied m times. This indicates pixel-level addition.

[0069] In the above scheme, the widely used adaptive momentum gradient descent algorithm was selected to construct the deep network optimizer. During training, the learning rate was set to 0.0005~0.005, the momentum was set to 0.9, and the widely adopted L1 cost function was used. The network was trained on internationally recognized datasets and a self-constructed tunnel entrance image dataset to obtain the optimal parameter values. The experimental environment used was an Intel processor, 16GB of RAM, 2 NVIDIA GetForce GTX 1050 Ti GPUs, and a PyTorch deep learning architecture for training and testing. Before training, the image set needs to be preprocessed. To avoid loss of generality, a 256x256 image patch was randomly selected from each training image. These randomly selected image patches were then used to construct training samples, and the deep learning model designed in this invention was used for training to obtain the optimal network parameters. During testing, the trained enhanced network model was used to obtain high-contrast, clear, and noise-free tunnel entrance enhanced images. The time complexity of the tests is also low; for example, the time taken to enhance an 800x800 resolution image is 0.05 seconds.

[0070] The above provides a detailed description of a tunnel entrance image enhancement method based on group attention kernel feedback perception provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the technical solutions and core ideas of this application. Those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for enhancing tunnel entrance images based on group attention kernel feedback perception, characterized in that, Includes the following steps: Build the encoding module; Construct a kernel feedback sensing network; Construct a decoding network; Construct a tunnel entrance image enhancement network with group attention kernel feedback perception; Construct a cost function to obtain the optimal parameters of the deep network, and obtain the enhanced image of the tunnel entrance based on the optimal network parameters; The constructed encoding module is composed of two cascaded group attention convolutional modules; wherein, each group attention convolutional module is composed of two convolutional blocks and one cascaded group attention module; each convolutional block consists of one convolutional layer with a kernel width of 3, one batch processing layer and one ReLU activation function; A kernel feedback perception network is constructed to continuously adapt and combine the fuzzy kernels of the convolutional layers with the kernel feedback perception information of the next layer. The kernel feedback perception network mainly consists of two branches: a kernel adaptive correction network and a group attention convolutional deep feature extraction network. The kernel feedback perception network first obtains a noisy feature map through a dimension stretching method, followed by two convolutional blocks and one 3x3 convolutional layer. The group attention convolutional deep feature extraction network first passes through a group attention convolutional module, followed by a deep feature extraction block. After the two parallel branches are completed, the visual features obtained from the two parallel branches are concatenated together, and then the concatenated visual features are reduced to the same channel dimension as the initial features through a 1x1 convolutional layer. Then, through short links, the visual features of the previous layer and the visual features that have undergone a series of module operations are added pixel-wise to obtain the input features of the kernel feedback perception network of the next layer. A tunnel entrance image enhancement network with kernel feedback perception is constructed. The network mainly consists of a feature encoding module, a kernel feedback module, and a feature decoding module. The enhancement network first extracts visual features through the feature encoding module, then performs m kernel self-feedback feature processing, and finally performs feature decoding to obtain high-frequency information of the visual features. Finally, through short links, the contour information of the input features is combined with the high-frequency information to form a tunnel entrance image with satisfactory visual quality, providing reliable visual information for further image understanding.

2. The tunnel entrance image enhancement method based on group attention kernel feedback perception according to claim 1, characterized in that: The constructed decoding network consists of two convolutional layers with edge padding.

3. The tunnel entrance image enhancement method based on group attention kernel feedback perception according to claim 1, characterized in that: The deep network was trained using the widely adopted adaptive momentum gradient descent algorithm network optimizer. During training, the learner rate was set to 0.0005~0.005, the momentum was set to 0.9, and the cost function used was the L1 cost function of the first paradigm. The network was effectively trained on internationally recognized datasets and a self-constructed tunnel entrance image dataset to obtain the optimal parameters for the augmented network.

Images