Mine low-illumination image enhancement method and system, electronic device and storage medium

Abstract

The application relates to a mine low-illumination image enhancement method and system, electronic equipment and a storage medium, the enhancement method comprising the following steps: acquiring two images with the same shooting angle under different brightness, namely a normal image and a low-illumination image; performing illumination decomposition operation on the normal image and the low-illumination image to obtain a first illumination component image, a second illumination component image, a first reflection component image and a second reflection component image; performing illumination adjustment processing on the second illumination component image by using the first illumination component image to obtain an illumination image; performing reflection reconstruction processing on the second reflection component image by using the first reflection component image to obtain a reflection image; and performing image fusion on the illumination image and the reflection image to obtain an enhanced image. The network structure used in the enhancement method can reduce data operation amount, improve operation speed, meanwhile, the enhancement effect and precision can be ensured, and a user can directly view the processed image to know the scene under the mine.

Description

Technical Field

[0001] This application relates to the field of image processing, and in particular to methods, systems, electronic devices, and storage media for enhancing low-light images in mines. Background Technology

[0002] Image enhancement technology is one of the fundamental aspects of digital image processing. Its purpose is to highlight certain information in an image according to specific needs, while weakening or removing unwanted information, making the processed image more suitable for human visual characteristics or machine recognition systems for a particular application. In practical applications, the original images acquired by our systems often fail to meet our needs. For example, in harsh working environments with complex geological conditions, such as underground mines, the quality of the original images acquired by the system is low due to noise, lighting, and other factors. Therefore, some processing is necessary to extract the information of interest.

[0003] Existing methods for enhancing mine images include: histogram equalization-based methods, homomorphic filtering-based methods, Retinex-based methods, and methods combining Retinex theory with deep neural networks. However, these methods have unsatisfactory enhancement effects or slow operation speeds, resulting in poor practicality. Summary of the Invention

[0004] The technical problem to be solved by this application is that existing mine image enhancement methods have unsatisfactory enhancement effects, or are slow in operation and lack practicality.

[0005] In order to solve the above problems, or at least partially solve the above technical problems, this application provides a method, system, electronic device and storage medium for enhancing low-light images in mines.

[0006] In a first aspect, the present invention provides a method for enhancing low-light images in mines, the specific steps of which are as follows:

[0007] Two images were captured from the same shooting angle under different brightness levels, namely a normal image and a low-light image;

[0008] Illumination decomposition is performed on both normal images and low-light images to obtain a first illumination component image, a second illumination component image, a first reflection component image, and a second reflection component image. The illumination decomposition operation includes inputting a normal image and a low-light image, and after the normal image and the low-light image are processed by the convolution module and the residual feature processing module of the illumination decomposition network, the first illumination component image and the second illumination component image are output through the activation function module.

[0009] The second illumination component image is input into the illumination adjustment network module;

[0010] The illumination adjustment network module operates through the convolution module and the residual feature processing module, and then outputs the illumination adjustment image through the activation function module.

[0011] The first illumination component image is merged with the illumination adjustment image to obtain the illumination image;

[0012] The second reflection component image is input into a lightweight depth-separable convolution module;

[0013] The reflection reconstruction image is obtained by operating the depthwise separable convolutional units of the lightweight depthwise separable convolutional module with the CA attention mechanism.

[0014] The first reflection component image is denoised and then merged with the reconstructed reflection image to output the reflection image.

[0015] The illumination image and the reflection image are fused to obtain an enhanced image.

[0016] Preferably, the illumination decomposition operation performed on both normal images and low-light images further includes the following steps:

[0017] Normal images and low-light images are input into a reflectance decomposition network for reflectance decomposition processing to obtain a first reflectance component image and a second reflectance component image.

[0018] The loss function value is obtained between the first illumination component image and the second illumination component image;

[0019] The loss function value is obtained between the first reflection component image and the second reflection component image.

[0020] Preferably, the operation of the convolution module and the residual feature processing module of the illumination decomposition network specifically includes the following steps:

[0021] Normal images and low-light images are simultaneously input into the feature processing module;

[0022] Normal images and low-light images are processed sequentially through three convolutional channels and a max pooling channel to obtain feature maps of four normal images and four low-light images, respectively.

[0023] The feature maps of four normal images are fused with the normal image to obtain the first illumination component image;

[0024] The feature maps of the four low-light images are fused with the low-light image to obtain the second illumination component image.

[0025] Preferably, the depthwise separable convolution operation and CA attention mechanism operation after passing through the lightweight depthwise separable convolution module specifically include the following steps:

[0026] Perform a normal convolution operation on each point of the second illumination component image to obtain the feature map after point convolution;

[0027] Each channel of the convolutional feature map is convolved to obtain a channel convolutional feature map.

[0028] The convolutional feature maps of all channels are superimposed to obtain the output feature map;

[0029] The output feature map is decomposed into two-dimensional features, and global average pooling is performed on the two dimensions to obtain a two-dimensional feature map.

[0030] The two-dimensional feature maps are fused to obtain the reflection reconstruction image.

[0031] In a second aspect, the present invention provides a low-light image enhancement system for mines, comprising:

[0032] Image acquisition unit: Acquires two images from the same shooting angle under different brightness levels, namely a normal image and a low-light image;

[0033] Illumination decomposition processing unit: Performs illumination decomposition operation on both normal images and low-light images to obtain a first illumination component image, a second illumination component image, a first reflection component image, and a second reflection component image;

[0034] Illumination adjustment unit: The first illumination component image performs illumination adjustment processing on the second illumination component image to obtain an illumination image;

[0035] Reflection Reconstruction Unit: The first reflection component image is used to reconstruct the second reflection component image to obtain a reflection image;

[0036] Image fusion unit: The illumination image and the reflection image are fused to obtain an enhanced image.

[0037] Thirdly, the present invention provides an electronic device, which includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0038] Memory, used to store computer programs;

[0039] When the processor executes the program stored in the memory, it implements the steps of the above-described method for enhancing low-light images in mines.

[0040] Fourthly, the present invention provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the mine low-light image enhancement method described above.

[0041] The technical solutions provided in this application have the following advantages compared with the prior art:

[0042] Two images with different brightness levels are acquired from the same shooting angle, resulting in a normal image and a low-light image. Both images undergo illumination decomposition processing. The normal image is processed to obtain a first illumination component image and a first reflection component image. The low-light image is processed to obtain a second illumination component image and a second reflection component image. The second illumination component image undergoes illumination adjustment processing, and the second reflection component image undergoes reflection reconstruction processing. Based on the first reflection component image and the first illumination component image, the second reflection component image and the second illumination component image can be corrected to enhance the image.

[0043] By enhancing low-light images to restore them to normal image quality, the image undergoes illumination decomposition. The Inception_ResNet illumination decomposition network combines the Inception module and the Residual module. This module fuses shallow and deep features, reducing information loss during illumination decomposition. The reflection reconstruction network employs a lightweight depthwise separable convolution module, which combines depthwise separable convolution with a CA attention mechanism. Depthwise separable convolution significantly reduces the number of network parameters by calculating for each channel, while the CA attention mechanism not only focuses on channel information but also has good spatial information extraction capabilities, ensuring enhancement effects while reducing computational load.

[0044] Furthermore, the network structure used in the enhancement method can reduce the amount of data computation and improve the running speed. At the same time, it can also ensure the enhancement effect and accuracy. Users can directly view the processed image and learn about the scene in the mine.

[0045] Furthermore, due to the use of a lightweight depth-separable convolution module, it can be used on portable electronic devices such as mobile phones, allowing for direct on-site processing and improving practicality. Attached Figure Description

[0046] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A schematic flowchart of a low-light image enhancement method for mines provided in the first embodiment of this application;

[0049] Figure 2 A detailed flowchart of step S2 of a low-light image enhancement method for mines provided in the first embodiment of this application;

[0050] Figure 3 A schematic diagram illustrating the specific process of step S21 in a low-light image enhancement method for mines provided in the first embodiment of this application;

[0051] Figure 4 A detailed flowchart illustrating step S212 of a low-light image enhancement method for mines provided in the first embodiment of this application;

[0052] Figure 5 A schematic diagram illustrating the specific process of step S3 in a low-light image enhancement method for mines provided in the first embodiment of this application;

[0053] Figure 6 A schematic diagram illustrating the specific process of step S4 in a low-light image enhancement method for mines provided in the first embodiment of this application;

[0054] Figure 7 A detailed flowchart of step S42 of a low-light image enhancement method for mines provided in the first embodiment of this application;

[0055] Figure 8 A block diagram of a low-light image enhancement system for mines provided in the second embodiment of this application;

[0056] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0057] Explanation of reference numerals in the attached figures:

[0058] 100. Low-light image enhancement system for mines; 10. Image acquisition unit; 20. Illumination decomposition and processing unit; 30. Illumination adjustment unit; 40. Reflection reconstruction unit; 50. Image fusion unit; 1. Electronic equipment; 11. Memory; 12. Processor. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0060] Firstly, see Figure 1 This invention provides a method for enhancing low-light images in mines, the specific steps of which are as follows:

[0061] Step S1: Acquire two images from the same shooting angle under different brightness levels, namely a normal image and a low-light image;

[0062] Step S2: Perform illumination decomposition on both the normal image and the low-light image to obtain the first illumination component image, the second illumination component image, the first reflection component image, and the second reflection component image;

[0063] Step S3: The first reflection component image is subjected to illumination adjustment processing on the first illumination component image to obtain an illumination adjustment image;

[0064] Step S4: The second reflection component image is subjected to reflection reconstruction processing on the second illumination component image to obtain the reflection reconstruction image;

[0065] Step S5: The illumination adjustment image and the reflection reconstruction image are fused to obtain the enhanced image.

[0066] Specifically, in step S1, two images with different brightness are captured at the same shooting angle: one taken under normal brightness and the other under low light conditions. Through subsequent processing, the images under normal brightness and low light conditions are processed to obtain an image of the real scene.

[0067] Specifically, in step S2, the Retinex model, constructed based on the illumination-reflection principle, represents the image as the product of the reflected light component and the illumination component, which can be expressed by the following formula: S(x,y)=R(x,y) I(x,y); where I(x,y) represents the reflection component, indicating the reflection characteristics of the object's surface; R(x,y) represents the illumination component, indicating the characteristics of ambient light; and S(x,y) is the observed image. For illumination decomposition processing of the image, this example uses the Inception_ResNet network as the illumination decomposition network. The Inception_ResNet network is mainly composed of a ResNet network and an Inception network. After performing illumination decomposition processing on the input normal image and low-light image using the ResNet network, the reflection component image is obtained. After performing illumination decomposition processing on the input normal image and low-light image using the Inception network, the illumination component image is obtained. Specifically, after illumination decomposition processing, the normal image yields a first reflection component image and a first illumination component image, and after illumination decomposition processing, the low-light image yields a second reflection component image and a second illumination component image.

[0068] Specifically, in step S3, the illumination adjustment processing is performed based on the illumination adjustment network. The illumination adjustment network adopts the Illumination_Adjustment network, and its network structure is similar to that of the illumination decomposition network. The illumination adjustment processing is performed on the first illumination component image to correct the illumination components of the first illumination component image. In addition, the first reflection component image is used as the basis for illumination adjustment processing to obtain the loss function value between the two, so as to facilitate subsequent machine learning.

[0069] Specifically, in step S4, the reflection reconstruction processing is performed based on a reflection reconstruction processing network, employing a lightweight depthwise separable convolutional module. This lightweight depthwise separable convolutional module includes depthwise separable convolutional units and a CA (Coordinate Attention) attention mechanism unit, which can significantly reduce the number of model parameters and improve model performance. Furthermore, the second reflection component image serves as the basis for reflection reconstruction processing, and the loss function value between the two is obtained to facilitate subsequent machine learning.

[0070] Specifically, in step S5, the obtained illumination adjustment image and reflection reconstruction image are fused to improve the utilization rate of image information, reduce information redundancy and uncertainty, restore low-light images, and correct normal images to obtain images of real scenes.

[0071] Specifically, two images with different brightness levels are acquired from the same shooting angle, resulting in a normal image and a low-light image. Both the normal and low-light images undergo illumination decomposition processing. After processing, the normal image yields a first illumination component image and a first reflection component image. After processing, the low-light image yields a second illumination component image and a second reflection component image. The second illumination component image undergoes illumination adjustment processing, and the second reflection component image undergoes reflection reconstruction processing. Based on the first reflection component image and the first illumination component image, the second reflection component image and the second illumination component image can be corrected to achieve image enhancement. The current network structure is a training method for enhancing low-light images to normal brightness images, which facilitates subsequent low-light image enhancement processing and improves the device's enhancement processing speed.

[0072] By enhancing low-light images to restore them to normal image quality, the image undergoes illumination decomposition. The Inception_ResNet illumination decomposition network combines the Inception module and the Residual module. This module fuses shallow and deep features, reducing information loss during illumination decomposition. The reflection reconstruction network employs a lightweight depthwise separable convolution module, which combines depthwise separable convolution with a CA attention mechanism. The CA attention mechanism not only focuses on channel information but also has good extraction capabilities for spatial information.

[0073] Furthermore, the network structure adopted can reduce the amount of data processing and improve the running speed. At the same time, it can also ensure the enhancement effect and accuracy. Users can directly view the processed images and learn about the scene in the mine.

[0074] Furthermore, due to its lightweight, depth-separable convolution module, it can be used on portable electronic devices such as mobile phones, allowing for direct on-site processing and improving practicality.

[0075] See Figure 2 Step S2 specifically includes the following steps:

[0076] Step S21: Input the normal image and the low-light image into the illumination decomposition network for illumination decomposition processing to obtain the first illumination component image and the second illumination component image;

[0077] Step S22: Input the normal image and the low-light image into the reflectance decomposition network for reflectance decomposition processing to obtain the first reflectance component image and the second reflectance component image;

[0078] Step S23: Obtain the loss function value between the first illumination component image and the second illumination component image;

[0079] Step S24: Obtain the loss function value between the first reflection component image and the second reflection component image.

[0080] Specifically, the reflection decomposition network uses a ResNet network, and the illumination decomposition network uses an Incpetion network. The reflection decomposition network can decompose the reflection components of normal and low-light images, while the illumination decomposition network can decompose the illumination components of normal and low-light images. The combination of Incpetion and ResNet networks requires less data and can effectively improve the running speed. The Incpetion network can perform feature processing on layers, enhancing the ability to extract deep features. In addition, the Incpetion_ResNet network can effectively reduce the computational cost of the network and speed up the training process, thus contributing to the improvement of network performance. Obtaining the loss function values ​​between the reflection and illumination components allows us to understand the current training status of the Incpetion_ResNet network, thereby evaluating the training results and facilitating subsequent adjustments.

[0081] It is understandable that the Inception-ResNet illumination decomposition network is composed of Inception and ResNet. Inception-ResNet combines the advantages of Inception and ResNet, maintaining the high accuracy of Inception while enabling the network to reach a deeper number of layers, thereby providing more powerful representation and judgment capabilities. Furthermore, Inception-ResNet can achieve network balance while maintaining high accuracy, making the training of each layer more stable and alleviating the overfitting phenomenon that is very easy to occur in deep neural networks.

[0082] Specifically, the reflection decomposition process in step S21 and the illumination decomposition process in step S22 are performed simultaneously. Both the reflection decomposition network and the illumination decomposition network perform illumination decomposition, aiming to output the illumination and reflection components of the normal and low-light images for subsequent processing. The order of their processing does not affect the illumination decomposition process. Furthermore, the illumination decomposition network and the reflection decomposition network have the same network structure and share weights.

[0083] See Figure 3 Step S21 specifically includes the following steps:

[0084] Step S211: Input normal image and low-light image;

[0085] Step S212: After the normal image and the low-light image are processed by the convolution module and the residual feature processing module of the illumination decomposition network, the first illumination component image and the second illumination component image are output through the activation function module.

[0086] Specifically, the illumination decomposition network includes a convolution module and a residual feature processing module. The input image is first subjected to a convolution operation with a kernel of 3, then processed by the residual feature processing module 8 times, and then subjected to another convolution operation with a kernel of 3. Finally, the decomposed image is output through the Leaky_ReLU activation function. The residual feature processing module is Ingress_ResNet_Block, which can effectively reduce the computational load of the network and speed up the network training, thus promoting the improvement of network performance.

[0087] See Figure 4 Step S212 specifically includes the following steps:

[0088] Step S2121: Input the normal image and the low-light image into the feature processing module simultaneously;

[0089] Step S2122: The normal image and the low-light image are processed by three convolution channels and a max pooling channel in sequence to obtain feature maps of four normal images and four low-light images respectively.

[0090] Step S2123: The feature maps of the four normal images are fused with the normal image to obtain the first illumination component image;

[0091] Step S2124: The feature maps of the four low-light images are fused with the low-light image to obtain the second illumination component image.

[0092] Specifically, the feature processing module includes three convolutional channels and a max-pooling channel. The three convolutional channels perform 1×1 conv, 1×1 conv + 3×3 conv, and 1×1 conv + 5×5 conv operations, respectively. The pooling channel performs 3×3 maxpooling + 1×1 conv operations. After the image is input, four feature maps with different numbers of channels but the same image size are obtained. The four feature maps are fused with the input image to obtain the corresponding illumination component images. The normal image undergoes the operation of the feature processing module to obtain the first illumination component image, and the low-light image undergoes the operation of the feature processing module to obtain the second illumination component image. The feature processing module combines different convolutional layers in parallel to efficiently expand the depth and width of the network. While maintaining the high accuracy of Inception, it enables the network to reach a deeper number of layers, thereby providing more powerful expression and judgment capabilities.

[0093] See Figure 5 Step S3 specifically includes the following steps:

[0094] Step S31: Input the second illumination component image into the illumination adjustment network module;

[0095] Step S32: The illumination adjustment network module operates through the convolution module and the residual feature processing module, and then outputs the illumination adjustment image through the activation function module;

[0096] Step S33: The first illumination component image and the illumination adjustment image are merged to obtain the illumination image.

[0097] Specifically, the illumination adjustment network module is similar in structure to the illumination decomposition network module. The residual feature processing module of the illumination adjustment network module has four fewer feature processing modules than the residual feature processing module of the illumination decomposition network module. It corrects the illumination component image, thereby outputting an image unaffected by illumination factors. The convolution module performs convolution operations with a kernel of 3. The residual feature processing module includes three convolution channels and a max pooling channel. The three convolution channels perform 1×1 conv, 1×1 conv + 3×3 conv, and 1×1 conv + 5×5 conv operations, respectively. The pooling channel performs 3×3 maxpooling + 1×1 conv operations. In the illumination adjustment network module, the image first undergoes convolution operations in the convolution module, then passes through the residual feature processing module to obtain four feature images, which are then fused with the input image to form a single output image. The convolution module then performs convolution operations again. Finally, the Leaky_ReLU activation function outputs the illumination adjustment image. The first illumination component image and the illumination adjustment image are fused to obtain the illumination image, which is an enhanced image of the illumination component of the low-light image.

[0098] In the illumination adjustment network module, image enhancement processing is performed on the illumination component image of the low-light image to obtain the illumination adjustment image. The illumination component image of the normal image is denoised. The illumination adjustment image and the illumination component image of the normal image are merged to obtain the enhanced image of the illumination component of the low-light image. This enhances the illumination component image of the low-light image. The value of the loss function between the illumination adjustment image and the first illumination component image is obtained. The current network training status can be known through the loss function value, and the input value and network parameters can be adjusted according to the current loss function value.

[0099] See Figure 6 Step S4 specifically includes the following steps:

[0100] Step S41: Input the second reflection component image into a lightweight depthwise separable convolutional module;

[0101] Step S42: The reflection reconstruction image is obtained by operating the depthwise separable convolutional units of the lightweight depthwise separable convolutional module with the CA attention mechanism;

[0102] Step S43: The first reflection component image is denoised and merged with the reflection reconstruction image to output the reflection image.

[0103] Specifically, the reflection reconstruction processing is based on a reflection reconstruction network, a lightweight depthwise separable convolutional module that includes depthwise separable convolutional units and a CA attention mechanism unit. This lightweight depthwise separable convolutional module significantly reduces the number of model parameters, making it suitable for deployment on mobile devices for deep learning tasks. This is primarily because it uses depthwise separable convolution instead of ordinary convolution, resulting in a significant performance improvement compared to previous networks. Furthermore, combined with the CA attention mechanism, feature maps can be acquired from both the width and height dimensions of the image, achieving the acquisition of remote dependencies and preservation of precise target location information from two dimensions, greatly improving model accuracy. By inputting the second reflection component image, reflection reconstruction processing is performed on the reflection component image of the low-light image, resulting in a reconstructed reflection image. After denoising the first reflection component image, the second reflection image and the reconstructed reflection image are merged to output the final reflection image. The loss function value between the reconstructed reflection image and the first reflection component image is obtained, allowing the user to understand the network processing status. Simultaneously, the input values ​​and network parameters can be adjusted based on the current loss function value.

[0104] See Figure 7 Step S42 specifically includes the following steps:

[0105] Step S421: Perform a normal convolution operation on each point of the second illumination component image to obtain the feature map after point convolution;

[0106] Step S422: Perform a convolution operation on each channel of the convolutional feature map to obtain the channel convolutional feature map;

[0107] Step S423: Stack all the channel convolutional feature maps together to obtain the output feature map;

[0108] Step S424: Decompose the output feature map into two-dimensional features, and perform global average pooling operation on the two dimensions to obtain two-dimensional feature maps;

[0109] Step S425: Fuse the two dimensional feature maps to obtain the reflection reconstruction image.

[0110] Specifically, the depthwise separable convolutional unit operation first performs pointwise convolution, and then performs depthwise convolution on the feature map. When performing depthwise convolution on the feature map, the image is decomposed into two different dimensions of features according to two different dimensions. After pooling, the processed feature maps are fused to obtain the inverse reconstructed image.

[0111] In step S421, a normal 1×1 convolution operation is performed on the input image to obtain a point convolution feature map. In step S422, a 3×3 convolution operation is performed on each channel of the obtained point convolution feature map to obtain a channel convolution feature map. In step S423, all channel convolution feature maps are superimposed to obtain an output feature map. In step S424, the output feature map is decomposed according to the width and height dimensions, and global average pooling is performed on the two dimensions to obtain feature maps in the height and width dimensions. In step S425, the feature maps in the two dimensions are fused to output the reflection reconstruction image.

[0112] Secondly, see Figure 8 The present invention provides a low-light image enhancement system 100 for mines, comprising:

[0113] Image acquisition unit 10: Acquires two images from the same shooting angle under different brightness levels, namely a normal image and a low-light image;

[0114] Illumination decomposition processing unit 20: Performs illumination decomposition operation on both normal images and low-illuminance images to obtain a first illumination component image, a second illumination component image, a first reflection component image and a second reflection component image;

[0115] Illumination adjustment unit 30: The first illumination component image performs illumination adjustment processing on the second illumination component image to obtain an illumination image;

[0116] Reflection reconstruction unit 40: The first reflection component image performs reflection reconstruction processing on the second reflection component image to obtain a reflection image;

[0117] Image fusion unit 50: The illumination adjustment image and the reflection reconstruction image are fused to obtain an enhanced image.

[0118] Specifically, the image acquisition unit 10 acquires images of normal illumination and low illumination at the same angle to facilitate subsequent algorithms in correcting the low illumination image and bringing it closer to the normal illumination image. The illumination decomposition processing unit 20 performs illumination decomposition operations on the normal and low illumination images, decomposing the images into illumination component images and reflection component images according to the Retinex model for subsequent processing. The illumination adjustment unit 30 performs illumination adjustment processing on the first illumination component image based on the illumination adjustment network, correcting the illumination component image to adjust it. The reflection reconstruction unit 40 performs reflection reconstruction processing on the second reflection component image based on the reflection reconstruction processing network, adjusting the second reflection component image to restore its normal brightness. The reflection reconstruction processing network uses a lightweight depthwise separable convolutional module, including depthwise separable convolutional units and CA (Coordinate Attention) attention mechanism units, which can greatly reduce the number of model parameters and improve model performance. In the image fusion unit 50, the illumination adjustment image obtained by the illumination adjustment unit 30 and the reflection reconstruction image obtained by the reflection reconstruction unit 40 are fused to adjust the low-light image to a normal image, so that the low-light image can clearly display the details of the normal scene.

[0119] It is understandable that by enhancing low-light images to restore them to normal image quality, the image undergoes illumination decomposition. The Inception_ResNet illumination decomposition network combines the Inception module and the Residual module. This module fuses shallow and deep features, reducing information loss during illumination decomposition. The reflection reconstruction network employs a lightweight depthwise separable convolution module, which combines depthwise separable convolution with a CA attention mechanism. The CA attention mechanism not only focuses on channel information but also has good spatial information extraction capabilities. Furthermore, the network structure reduces computational load and improves running speed while ensuring enhancement effect and accuracy. Users can directly view the processed image to understand the scene in the mine. Moreover, due to the lightweight depthwise separable convolution module, it can be used on portable electronic devices such as mobile phones, allowing for direct on-site processing and improving practicality.

[0120] The reflection reconstruction unit 40 specifically includes the following steps:

[0121] Image input unit: Lightweight depthwise separable convolution module for inputting the second reflection component image;

[0122] Image reconstruction acquisition unit: The reflection reconstruction image is obtained by operating the depthwise separable convolution unit of the lightweight depthwise separable convolution module and the CA attention mechanism;

[0123] Reflection image output unit: The first reflection component image is denoised and merged with the reflection reconstruction image to output the reflection image.

[0124] Specifically, the second illumination component image is input into the reflection reconstruction network for processing. After passing through separable convolutional units and CA attention mechanism operations, a reflection reconstruction image is obtained. In addition, a loss function operation is performed with the second reflection component to obtain the error between the current reflection reconstruction image and the second reflection component image from the loss function value, which is helpful for subsequent network training.

[0125] It is understandable that reflection reconstruction processing is based on a reflection reconstruction network, which is a lightweight deep separable convolutional module. It includes deep separable convolutional units and CA attention mechanism units. The lightweight deep separable convolutional module greatly reduces the number of model parameters and can be deployed on mobile devices to implement deep learning tasks. This is mainly because it uses deep separable convolution to replace ordinary convolution, and its model performance is significantly improved compared to previous networks. In addition, combined with the CA attention mechanism, feature maps can be obtained from both the width and height dimensions of the image, so as to obtain remote dependencies and retain accurate target location information from two dimensions, which greatly improves the accuracy of the model.

[0126] The reconstructed image acquisition unit specifically includes the following steps:

[0127] Point convolution processing unit: Performs ordinary convolution operation on each point of the second illumination component image to obtain the feature map after point convolution;

[0128] Channel convolution processing unit: Each channel of the convolved feature map is convolved to obtain the channel convolution feature map;

[0129] Feature map output unit: Superimposes all channel convolutional feature maps to obtain the output feature map;

[0130] Feature map decomposition unit: Decomposes the output feature map into two-dimensional features, performs global average pooling operation on the two dimensions, and obtains two-dimensional feature maps;

[0131] Feature map fusion unit: fuses two-dimensional feature maps to obtain a reflection reconstruction image.

[0132] Specifically, the depthwise separable convolutional unit operation first performs pointwise convolution, and then performs depthwise convolution on the feature map. When performing depthwise convolution on the feature map, the image is decomposed into two different dimensions of features according to two different dimensions. After pooling, the processed feature maps are fused to obtain the inverse reconstructed image.

[0133] It is understandable that in a lightweight depthwise separable convolution module, the input image can be processed through the operations of the depthwise separable convolution unit and the CA attention mechanism unit, adjusting the image to a normal illumination image. At the same time, due to the characteristics of depthwise separable convolution and CA attention mechanism, the amount of data processed can be greatly reduced. Meanwhile, the low-light image enhancement effect is good, ensuring the accuracy of the output image.

[0134] Thirdly, see Figure 9 The present invention provides an electronic device 1, which includes a processor 12, a communication interface, a memory 11 and a communication bus, wherein the processor 12, the communication interface and the memory 11 communicate with each other through the communication bus;

[0135] Memory 11 is used to store computer programs;

[0136] When the processor 12 executes the program stored in the memory 11, it implements the steps of the low-light image enhancement method for mines mentioned in the first aspect.

[0137] Specifically, the processor 12 runs the program stored in the memory 11 to implement the image enhancement method, which can enhance low-light images to restore the details of normal scenes and make it easier for users to view the details of the mine environment.

[0138] The low-light image enhancement method for mines employs the Inception_ResNet network and a lightweight depthwise separable convolutional module to decompose the image into reflection and illumination components. The two components of the low-light image are then processed by an illumination adjustment network and a reflection reconstruction network, and the processed images are fused to obtain the enhanced low-light image. This process restores the low-light image to a normal scene. Simultaneously, the loss function value is obtained by comparing the processed low-light image with the normal image. The error between the processed and normal images is used to determine whether the processing method is realistic, thereby enhancing network training and ensuring the enhancement effect of low-light images. Furthermore, the Inception_ResNet network and the lightweight depthwise separable convolutional module require less data processing, thus improving the running speed.

[0139] Fourthly, the present invention provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the mine low-light image enhancement method described above.

[0140] Specifically, the storage medium stores a computer program to execute and implement the above-mentioned enhancement method. This method ensures the enhancement effect of low-light images. At the same time, when the computer program is running, the amount of data processed can be greatly reduced, the execution speed can be improved, and the results can be obtained more quickly.

[0141] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0142] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0143] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0144] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0145] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0146] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. The illustrative expressions of the above terms in this specification should not be construed as necessarily referring to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0147] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Since these modifications and variations fall within the scope of the claims and their equivalents, this invention also intends to include these modifications and variations.

[0148] The above description describes specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for enhancing low-light images in mines, characterized in that, The specific steps are as follows: Two images were captured from the same shooting angle under different brightness levels, namely a normal image and a low-light image; Illumination decomposition is performed on both normal images and low-light images to obtain a first illumination component image, a second illumination component image, a first reflection component image, and a second reflection component image. The illumination decomposition operation includes inputting a normal image and a low-light image, and after the normal image and the low-light image are processed by the convolution module and the residual feature processing module of the illumination decomposition network, the first illumination component image and the second illumination component image are output through the activation function module. The second illumination component image is input into the illumination adjustment network module; The illumination adjustment network module operates through the convolution module and the residual feature processing module, and then outputs the illumination adjustment image through the activation function module. The first illumination component image is merged with the illumination adjustment image to obtain the illumination image; The second reflection component image is input into a lightweight depth-separable convolution module; The reflection reconstruction image is obtained by operating the depthwise separable convolutional units of the lightweight depthwise separable convolutional module with the CA attention mechanism. The first reflection component image is denoised and then merged with the reconstructed reflection image to output the reflection image. The illumination image and the reflection image are fused to obtain an enhanced image.

2. The method according to claim 1, characterized in that, The illumination decomposition operation performed on both normal and low-light images specifically includes the following steps: Normal images and low-light images are input into a reflectance decomposition network for reflectance decomposition processing to obtain a first reflectance component image and a second reflectance component image. The loss function value is obtained between the first illumination component image and the second illumination component image; The loss function value is obtained between the first reflection component image and the second reflection component image.

3. The method according to claim 1, characterized in that, The operation of the convolutional module and residual feature processing module of the illumination decomposition network specifically includes the following steps: Normal images and low-light images are simultaneously input into the feature processing module; Normal images and low-light images are processed sequentially through three convolutional channels and a max pooling channel to obtain feature maps of four normal images and four low-light images, respectively. The feature maps of four normal images are fused with the normal image to obtain the first illumination component image; The feature maps of the four low-light images are fused with the low-light image to obtain the second illumination component image.

4. The method according to claim 1, characterized in that, The depthwise separable convolution operation and CA attention mechanism operation via the lightweight depthwise separable convolution module specifically include the following steps: Perform a normal convolution operation on each point of the second illumination component image to obtain the feature map after point convolution; Each channel of the convolutional feature map is convolved to obtain a channel convolutional feature map. The convolutional feature maps of all channels are superimposed to obtain the output feature map; The output feature map is decomposed into two-dimensional features, and global average pooling is performed on the two dimensions to obtain a two-dimensional feature map. The two-dimensional feature maps are fused to obtain the reflection reconstruction image.

5. A low-light image enhancement system for mines, characterized in that, include: Image acquisition unit: Acquires two images from the same shooting angle under different brightness levels, namely a normal image and a low-light image; Illumination decomposition processing unit: Performs illumination decomposition operation on both normal image and low-light image to obtain first illumination component image, second illumination component image, first reflection component image and second reflection component image; wherein, the illumination decomposition operation includes, inputting normal image and low-light image, after the normal image and low-light image are processed by the convolution module and residual feature processing module of illumination decomposition network, the first illumination component image and second illumination component image are output through activation function module; Illumination Adjustment Unit: The second illumination component image is input into the illumination adjustment network module; after being processed by the convolution module and residual feature processing module of the illumination adjustment network module, the illumination adjustment image is output through the activation function module; the first illumination component image and the illumination adjustment image are merged to obtain the illumination image. Reflection Reconstruction Unit: The second reflection component image is input to a lightweight depthwise separable convolutional module; after operation by the depthwise separable convolutional unit and CA attention mechanism of the lightweight depthwise separable convolutional module, a reflection reconstruction image is obtained; the first reflection component image undergoes noise reduction processing and is merged with the reflection reconstruction image to output the reflection image. Image fusion unit: The illumination image and the reflection image are fused to obtain an enhanced image.

6. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the low-light image enhancement method for mines as described in any one of claims 1-4.

7. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the low-light image enhancement method for mines as described in any one of claims 1-4.

Images