Mine image enhancement method and device and electronic equipment

Abstract

The application provides a mine image enhancement method and device and electronic equipment. The method comprises the following steps: obtaining a mine image, performing transform decomposition on the mine image by using an enhancement model to obtain a first branch image and a second branch image; performing amplitude spectrum extraction on the first branch image to obtain an initial amplitude spectrum feature, and generating an amplitude attention map feature by using an amplitude attention mechanism; determining an image brightness enhancement feature based on the initial amplitude spectrum feature and the amplitude attention map feature; performing phase extraction on the second branch image to obtain an initial phase feature, and generating a phase attention map feature by using a phase attention mechanism; determining a pixel structure enhancement feature based on the initial phase feature and the phase attention map feature; determining an enhanced mine image based on the image brightness enhancement feature and the pixel structure enhancement feature, and outputting the enhanced mine image by using the enhancement model, thereby solving the technical problem of poor enhancement effect of the mine image in the prior art.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus and electronic device for enhancing mine images. Background Technology

[0002] Images from underground coal mines commonly suffer from uneven lighting, low contrast, loss of detail, and noise interference, severely impacting the accuracy of visual tasks such as target detection. Enhancing the quality of underground coal mine images has become a bottleneck for intelligent monitoring and target recognition in underground operations. Existing image enhancement methods can be categorized into three types. The first type is based on encoder-decoder architectures, but suffers from detail loss. The second type focuses on improving spatial convolutional structures, but has a limited receptive field and is prone to losing edge structural information. The third type focuses on using frequency domain information to guide spatial feature extraction, but easily introduces artifacts at structural boundaries. Existing image enhancement methods result in poor image enhancement effects for mine images. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide a method, apparatus and electronic device for enhancing mine images, so as to overcome all or part of the shortcomings of the prior art.

[0004] To achieve the above objectives, this application provides a method for enhancing mine images, comprising: acquiring a mine image and inputting the mine image into a pre-trained enhancement model; performing transform decomposition on the mine image through the enhancement model to obtain a first branch image and a second branch image; extracting the amplitude spectrum of the first branch image to obtain initial amplitude spectrum features; and generating amplitude attention map features corresponding to the first branch image using an amplitude attention mechanism; determining image brightness enhancement features corresponding to the first branch image based on the initial amplitude spectrum features and the amplitude attention map features; extracting the phase of the second branch image to obtain initial phase features; and generating phase attention map features corresponding to the second branch image using a phase attention mechanism; determining pixel structure enhancement features corresponding to the second branch image based on the initial phase features and the phase attention map features; determining an enhanced mine image based on the image brightness enhancement features and the pixel structure enhancement features; and outputting the enhanced mine image through the enhancement model.

[0005] Optionally, determining the enhanced mine image based on the image brightness enhancement features and the pixel structure enhancement features includes: transforming and stitching the image brightness enhancement features and the pixel structure enhancement features to obtain stitched features; using the encoder in the enhancement model to perform multi-level spatial feature extraction on the stitched features to obtain multiple enhanced stitched features; using the cascaded multi-scale dynamic serpentine convolution in the enhancement model to extract the local features corresponding to each enhanced stitched feature; using the multi-hole complex convolution with multiple void ratios in the enhancement model to model the multiple enhanced stitched features to obtain global features; and using the decoder in the enhancement model to perform image reconstruction on the global features and the local features corresponding to each enhanced stitched feature to obtain the enhanced mine image.

[0006] Optionally, generating the amplitude attention map features corresponding to the first branch image using the amplitude attention mechanism includes: performing a color space transformation on the first branch image to obtain a transformed first branch image; decomposing the transformed first branch image to obtain an illuminance map; extracting multi-scale luminance features from the illuminance map to obtain luminance features; extracting the amplitude spectrum from the luminance features to obtain the luminance distribution corresponding to the luminance features; and generating the amplitude attention map features corresponding to the first branch image based on the luminance distribution.

[0007] Optionally, generating the phase attention map features corresponding to the second branch image using the phase attention mechanism includes: performing a phase transformation on the second branch image to obtain a transformed second branch image; extracting multi-scale structural features from the transformed second branch image to obtain structural features; performing phase spectral extraction on the structural features to obtain structural contour features; and generating the phase attention map features corresponding to the second branch image based on the structural contour features.

[0008] Optionally, determining the image brightness enhancement feature corresponding to the first branch image based on the initial amplitude spectrum feature and the amplitude attention map feature includes: multiplying the initial amplitude spectrum feature and the amplitude attention map feature element-wise to obtain a first weighted feature; and concatenating the first weighted feature with the initial amplitude spectrum feature using residual concatenation to obtain the image brightness enhancement feature.

[0009] Optionally, determining the pixel structure enhancement features corresponding to the second branch image based on the initial phase features and the phase attention map features includes: multiplying the initial phase features and the phase attention map features element-wise to obtain a second weighted feature; and concatenating the second weighted feature with the initial phase features using residual concatenation to obtain the pixel structure enhancement features.

[0010] Optionally, the loss function used to train the augmented model is determined by the following formula: ,in, Represents the loss function. This indicates the magnitude loss guided by illuminance. For illuminance adaptive weighting coefficients, This represents the frequency-domain weighted cosine phase loss. These are frequency domain adaptive weighting coefficients. For the histogram loss in the spatial domain, These are spatial adaptive weighting coefficients. In order to perceive loss, For sensing adaptive weighting coefficients.

[0011] Optionally, the training data used to train the enhancement model is determined by the following method: acquiring multiple historical mine images, wherein the brightness of each historical mine image is less than or equal to a predetermined brightness; in response to determining that the number of the multiple historical mine images is less than a predetermined number, acquiring multiple normal images with brightness greater than the predetermined brightness; for each normal image, segmenting the normal image to obtain multiple superpixel regions; for each superpixel region corresponding to each normal image, randomly generating an attenuation coefficient for the superpixel region, and using Gaussian smoothing to attenuate the brightness of the superpixel region based on the attenuation coefficient; in response to determining that the brightness of the superpixel region after brightness attenuation is less than or equal to the predetermined brightness, using the superpixel region after brightness attenuation as a new image; in response to determining that the total number of the multiple historical mine images and all new images is greater than or equal to the predetermined number, using the multiple historical mine images and all new images together as training data for the enhancement model.

[0012] Based on the same inventive concept, this application also provides a mine image enhancement device, comprising: a decomposition module configured to acquire a mine image and input the mine image into a pre-trained enhancement model, and perform transform decomposition on the mine image through the enhancement model to obtain a first branch image and a second branch image; a first generation module configured to extract the amplitude spectrum of the first branch image to obtain initial amplitude spectrum features, and generate amplitude attention map features corresponding to the first branch image using an amplitude attention mechanism; a first determination module configured to determine image brightness enhancement features corresponding to the first branch image based on the initial amplitude spectrum features and the amplitude attention map features; a second generation module configured to extract the phase of the second branch image to obtain initial phase features, and generate phase attention map features corresponding to the second branch image using a phase attention mechanism; a second determination module configured to determine pixel structure enhancement features corresponding to the second branch image based on the initial phase features and the phase attention map features; and a third determination module configured to determine the enhanced mine image based on the image brightness enhancement features and the pixel structure enhancement features, and output the enhanced mine image through the enhancement model.

[0013] Based on the same inventive concept, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.

[0014] Based on the same inventive concept, this application also provides a non-transitory computer-readable storage medium that stores computer instructions for causing a computer to perform the method described above.

[0015] As can be seen from the above description, the mine image enhancement method, apparatus, and electronic device provided in this application include: acquiring a mine image and inputting the mine image into a pre-trained enhancement model; performing transform decomposition on the mine image through the enhancement model to obtain a first branch image and a second branch image; extracting the amplitude spectrum of the first branch image to obtain initial amplitude spectrum features; and generating amplitude attention map features corresponding to the first branch image using an amplitude attention mechanism; determining image brightness enhancement features corresponding to the first branch image based on the initial amplitude spectrum features and the amplitude attention map features; extracting the phase of the second branch image to obtain initial phase features; and generating phase attention map features corresponding to the second branch image using a phase attention mechanism; determining pixel structure enhancement features corresponding to the second branch image based on the initial phase features and the phase attention map features; determining the enhanced mine image based on the image brightness enhancement features and the pixel structure enhancement features; and outputting the enhanced mine image through the enhancement model. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic flowchart of a method for enhancing mine images according to an embodiment of this application;

[0018] Figure 2 This is a schematic diagram of the core network structure APSFNet of the enhanced model in an embodiment of this application;

[0019] Figure 3 This is a schematic diagram of the structure of the dual-domain fusion module (DDFModule) according to an embodiment of this application, wherein, Figure 3 (a) is a schematic diagram of the dual-domain fusion module. Figure 3 (b) is a schematic diagram of frequency domain branching. Figure 3 (c) is a schematic diagram of the airspace branch;

[0020] Figure 4 This is a schematic diagram of the Amplitude Attention Module (AAM) according to an embodiment of this application;

[0021] Figure 5 This is a schematic diagram of the phase attention module (PAM) according to an embodiment of this application;

[0022] Figure 6This is a schematic diagram of the structure of the multi-directional Gabor convolution module MGCM according to an embodiment of this application;

[0023] Figure 7 This is a subjective visual comparison image of a comparative experiment on the self-made dataset TG-ULLID in underground coal mines, as described in this application.

[0024] Figure 8 These are subjective visual comparison images of the comparative experiments on the LOL-v2 public dataset in the embodiments of this application;

[0025] Figure 9 The graph shows a comparative experiment of this application's embodiments on the public dataset DsLMF+helmet.

[0026] Figure 10 This is a schematic diagram of the structure of the mine image enhancement device according to an embodiment of this application;

[0027] Figure 11 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0029] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0030] As described in the background section, underground coal mine images commonly suffer from uneven illumination, low contrast, blurred targets, loss of detail, and noise interference due to factors such as low light intensity, point light sources, and high dust levels. These issues directly affect the accuracy of visual tasks such as personnel detection, safety helmet identification, and hazard source monitoring. Enhancing the quality of underground coal mine images has become a bottleneck problem for intelligent underground monitoring and target recognition. Therefore, researching image enhancement methods to solve the image quality problems of underground video surveillance is of great significance for improving the level of intelligent mine safety monitoring and achieving inherent safety.

[0031] With the rapid development of artificial intelligence technology, deep learning-based image enhancement methods have become a core technical approach and research hotspot in this field. Based on network structure evolution, existing image enhancement methods can be divided into two categories. The first category is based on encoder-decoder architectures, such as directly learning enhancement functions through end-to-end mapping and using stacked autoencoders to improve overall brightness, but this suffers from detail loss. To improve feature representation and computational efficiency, the second category focuses on improving spatial convolutional structures. For example, lightweight networks are built using depthwise separable convolutions; however, the receptive field is limited, and edge structure information is easily lost. In recent years, to better distinguish image structure from noise, researchers have begun to explore the fusion mechanism of the frequency and spatial domains, forming a third category of methods. Frequency domain information is used to guide spatial feature extraction, but artifacts are easily introduced at structural boundaries. Existing image enhancement methods result in poor image enhancement effects for mine images.

[0032] In view of this, embodiments of this application propose a method for enhancing mine images, referring to... Figure 1 This includes the following steps:

[0033] Step 101: Obtain a mine image and input the mine image into a pre-trained enhancement model. The enhancement model performs a transform decomposition on the mine image to obtain a first branch image and a second branch image.

[0034] In this step, due to factors such as low illumination, point light sources, and high dust levels within the mine, mine images generally suffer from uneven lighting, low contrast, blurred targets, loss of detail, and noise interference. Directly using mine images for identification would lead to low accuracy. To address these issues, this application utilizes a pre-trained enhancement model to enhance mine images, subsequently using the enhanced images for identification, significantly improving accuracy. First, mine images captured within the mine are acquired. These images are then input into the pre-trained enhancement model, which decomposes them into two branches using a Fast Fourier Transform (FFT). The first branch handles the brightness information, while the second branch handles the structural information. Utilizing the FFT reduces computational complexity and efficiently decomposes the mine images according to different dimensions, enabling targeted image enhancement from various dimensions.

[0035] Step 102: Extract the amplitude spectrum of the first branch image to obtain the initial amplitude spectrum features, and generate the amplitude attention map features corresponding to the first branch image using the amplitude attention mechanism.

[0036] In this step, amplitude spectrum extraction is performed on the first branch image to obtain initial amplitude spectrum features. Specifically, two 1×1 convolutions and the LeakyReLU activation function are used to process the initial amplitude spectrum, generating initial amplitude spectrum features to achieve the preliminary and accurate determination of the amplitude-related features of the first branch image. The amplitude spectrum features reflect the distribution patterns and energy characteristics of different frequency components in the first branch image. Using two 1×1 convolutions combined with the LeakyReLU activation function to process the initial amplitude spectrum enables efficient feature extraction and dimensionality reduction through linear combination and nonlinear activation between channels. The first layer of 1×1 convolution expands the number of channels to capture the local correlation of different frequency components in the amplitude spectrum, and the second layer further fuses features and controls the output dimension, reducing computational complexity. LeakyReLU avoids neuron inactivation by preserving negative information, enhancing the model's ability to model weak signals in the amplitude spectrum (such as structural edges or high-frequency noise under low light) while maintaining gradient fluidity to accelerate convergence; this lightweight structure can significantly improve the expressive power of features in mine image processing.

[0037] An amplitude attention mechanism is used to generate amplitude attention map features corresponding to the first branch image. These amplitude attention map features are feature representations generated by combining the amplitude spectrum with the attention mechanism. By using this mechanism to generate the amplitude attention map features corresponding to the first branch image, key brightness regions in the first branch image are dynamically focused on, and irrelevant interference is suppressed, thereby achieving the goal of further accurately determining the amplitude-related features of the first branch image.

[0038] Step 103: Based on the initial amplitude spectrum features and the amplitude attention map features, determine the image brightness enhancement features corresponding to the first branch image.

[0039] In this step, the initial amplitude spectrum features and amplitude attention map features are used to jointly determine the image brightness enhancement features corresponding to the first branch image. Combining the global illumination distribution (initial amplitude spectrum features) and local key regions (amplitude attention map features) to determine the image brightness enhancement features can ensure overall brightness coordination and avoid overexposure / underexposure through low-frequency global information, while using high-frequency local attention to accurately locate key regions such as dark areas and edges for dynamic enhancement. This effectively solves the problems of loss of details in global processing and illumination distortion caused by local processing in traditional methods, ultimately achieving a natural brightness, detail preservation, and strong anti-interference enhancement effect.

[0040] Step 104: Perform phase extraction on the second branch image to obtain initial phase features, and use the phase attention mechanism to generate the phase attention map features corresponding to the second branch image.

[0041] In this step, phase extraction is performed on the second branch image to obtain initial phase features. Specifically, this is achieved by processing the image with two 1×1 convolutional layers and the LeakyReLU activation function, thus generating initial phase features to accurately determine the phase-related features of the second branch image. These initial phase features reflect the spatial distribution and structural relationships of different frequency components in the second branch image. Phase extraction of the second branch image using two 1×1 convolutional layers and the LeakyReLU activation function efficiently captures spatial structural information (such as edges and texture directions) in the phase spectrum with lightweight computation. This avoids the computational burden of directly processing high-dimensional phase data and enhances feature representation through nonlinear activation. The initial phase features retain image details while possessing adaptability, providing structural complementarity for subsequent fusion with amplitude features, ultimately improving edge sharpness and overall consistency in brightness enhancement or restoration tasks.

[0042] Phase attention mechanisms are used to generate phase attention map features corresponding to the second branch image. These phase attention map features are feature representations generated by combining the phase spectrum and the attention mechanism. Generating phase attention map features for the second branch image using this mechanism can dynamically focus on key structural regions such as edges and contours in the image, while suppressing interference from noise or low-response regions. This is particularly suitable for tasks requiring high-fidelity structural restoration, such as low-light enhancement and super-resolution reconstruction, thereby achieving the goal of further accurately determining the phase-related features of the second branch image.

[0043] Step 105: Based on the initial phase features and the phase attention map features, determine the pixel structure enhancement features corresponding to the second branch image.

[0044] In this step, the pixel structure enhancement features corresponding to the second branch image are jointly determined using the initial phase features and the phase attention map features. By fusing the initial phase features (encoding the overall spatial structure information of the image) and the phase attention map features (dynamically focusing on key structural regions and suppressing interference), the pixel structure enhancement features are jointly determined. This approach can both utilize the global structural prior of the initial phase to avoid local enhancement distortion and accurately enhance details such as edges and textures through attention weighting. Ultimately, pixel-level structure enhancement features with both overall consistency and local sharpness are generated, significantly improving the structural fidelity and visual quality in image restoration tasks (such as low-light enhancement and super-resolution reconstruction).

[0045] It should be noted that the core network structure of the enhancement model of the coal mine underground image enhancement method of this application, which combines amplitude phase feature synergy and spatial frequency domain feature fusion (APSFNet), is as follows: Figure 2 As shown, it consists of two sub-networks: a Frequency Enhancement Network (FreqEnhanceNet) and a Spatial-Frequency Fusion Network (SpatFreqFusionNet). The Frequency Enhancement Network in the enhancement model is responsible for reconstructing the image in the frequency domain. An amplitude attention mechanism is generated by introducing an illumination map to improve the enhancement accuracy, and a phase attention mechanism is generated by using phase consistency priors to maintain structural integrity, so as to determine the image brightness enhancement features and pixel structure enhancement features respectively.

[0046] Step 106: Based on the image brightness enhancement features and the pixel structure enhancement features, determine the enhanced mine image, and output the enhanced mine image through the enhancement model.

[0047] In this step, brightness enhancement features restore the reasonable illumination distribution of the image through global or local brightness adjustments; pixel structure enhancement features accurately locate and enhance high-frequency structural information such as edges and textures, suppress noise interference, and ensure that key structural features are preserved. The fusion of these two features avoids structural loss during brightness enhancement and prevents brightness distortion during structure enhancement, ultimately generating a mine image with natural brightness, moderate contrast, and sharp details, ensuring effective enhancement of the mine image.

[0048] The above scheme involves acquiring a mine image and inputting it into a pre-trained enhancement model. The enhancement model performs transform decomposition on the mine image to obtain a first branch image and a second branch image. Amplitude spectrum extraction is performed on the first branch image to obtain initial amplitude spectrum features. An amplitude attention map feature corresponding to the first branch image is generated using an amplitude attention mechanism. Based on the initial amplitude spectrum features and the amplitude attention map feature, image brightness enhancement features corresponding to the first branch image are determined. Phase extraction is performed on the second branch image to obtain initial phase features. A phase attention map feature corresponding to the second branch image is generated using a phase attention mechanism. Based on the initial phase features and the phase attention map feature, pixel structure enhancement features corresponding to the second branch image are determined. Based on the image brightness enhancement features and the pixel structure enhancement features, an enhanced mine image is determined, and the enhanced mine image is output through the enhancement model.

[0049] In some embodiments, determining the enhanced mine image based on the image brightness enhancement features and the pixel structure enhancement features includes: transforming and stitching the image brightness enhancement features and the pixel structure enhancement features to obtain stitched features; using the encoder in the enhancement model to perform multi-level spatial feature extraction on the stitched features to obtain multiple enhanced stitched features; using the cascaded multi-scale dynamic serpentine convolution in the enhancement model to extract the local features corresponding to each enhanced stitched feature; using the multi-void complex convolution with multiple void ratios in the enhancement model to model the multiple enhanced stitched features to obtain global features; and using the decoder in the enhancement model to perform image reconstruction on the global features and the local features corresponding to each enhanced stitched feature to obtain the enhanced mine image.

[0050] In this embodiment, the image brightness enhancement features and pixel structure enhancement features are transformed and concatenated to obtain the concatenated features. Specifically, the image brightness enhancement features and pixel structure enhancement features are subjected to Inverse Fast Fourier Transform (IFFT) to obtain the concatenated features. Feature decoupling is achieved through frequency domain separation: the enhancement effects of brightness and structure are optimized separately in the frequency domain to avoid artifacts such as halos and noise caused by direct fusion in the spatial domain. Then, the optimized frequency domain features are losslessly converted back to the spatial domain by IFFT, finally obtaining a concatenated image with natural brightness transition and sharp details, which significantly improves the visual quality and algorithm robustness of image restoration in complex lighting scenes.

[0051] While FreqEnhanceNet effectively enhances the overall brightness of images, it still suffers from problems such as uneven illumination and noise amplification. To address these issues, this application proposes a SpatFreqFusionNet network based on an encoder-decoder structure, focusing on spatial domain reconstruction. Its core is the dual-domain fusion module DDFModule. This module employs a dual-branch architecture, fusing multi-scale depthwise separable convolutions and complex multi-scale dilated convolutions, fully utilizing their complementarity to collaboratively recover the structural and textural details of the image. The network extracts multi-level spatial features from the stitched features through the encoder, obtaining multiple enhanced stitched features. The DDFModule acts as a bridge between the encoder and decoder, with the specific structure as follows: Figure 3 As shown in (a), this module adopts a dual-branch parallel structure: the spatial branch extracts local features through cascaded multi-scale DSConv (cascaded multi-scale dynamic snake convolution), focusing on image detail restoration, edge sharpening, and texture preservation; the frequency domain branch models multiple enhanced stitching features using dilated complex convolution (Dilated ComplexConv) with multiple dilation rates to obtain global features, which are used for tone adjustment, exposure correction, and high-level semantic understanding. By effectively fusing local details and global information, this module can collaboratively optimize structural integrity and illumination distribution uniformity during image reconstruction, thereby improving visual quality while suppressing noise interference and uneven illumination. Using the decoder in the enhancement model, the global features and the local features corresponding to each enhanced stitching feature are used to restore the image, resulting in an enhanced mine image. The decoder is then used to fuse multi-resolution features to reconstruct the enhanced image. The stitching features are processed through the SpatFreqFusionNet network, eliminating the problems of uneven illumination and noise amplification, ensuring effective enhancement of the mine image.

[0052] The specific processing flow of the dual-domain fusion module (DDFModule) is as follows: Figure 3 As shown in (a), firstly, the spliced ​​features enhanced in the frequency domain are extracted using a 3×3 convolution, and then the resulting features are fed into the spatial domain branch and the frequency domain branch respectively. Figure 3 As shown in (c), the spatial branch uses multi-scale DSConv (depth-separable convolution) with kernel sizes of 1×1, 3×3, and 5×5 to extract and fuse multi-scale shallow features. The calculation formula is as follows:

[0053] Formula 1

[0054] in, Indicates local features, For depthwise separable convolutions with a kernel size of 1×1, This is a depthwise separable convolution with a kernel size of 3×3. For a depthwise separable convolution with a kernel size of 5×5, the formula for calculating DSConv for k×k (k=1, 3, or 5) is:

[0055] Formula 2

[0056] in, This indicates a depthwise separable convolution with a kernel size of k×k. This represents a depthwise convolution operation. For the input feature map, Represents the depthwise convolution kernel. Represents the pointwise convolution kernel. This represents a point convolution operation. This represents the concatenation operation performed along the channel dimension. Each scale branch, after convolution, sequentially employs GN and ReLU activation functions to synergistically improve model stability and enhance the non-linear expressive power of features. For example... Figure 3 As shown in (b), in the frequency domain branch, the input features are first transformed to the frequency domain using FFT. To overcome the limitations of real convolution in processing complex frequency domain signals and modeling phase information, complex convolution is used, and its calculation formula is: Formula 3

[0057] Where W=A+jB is the complex-valued convolution kernel, and z=x+jy is the complex-valued feature map. Let be the real part of the convolution kernel. The real part of the feature map, y is the imaginary part of the convolution kernel, y is the imaginary part of the feature map, and j is the imaginary unit. This indicates a convolution operation.

[0058] To achieve multi-scale feature fusion in the frequency domain, this application designs a dilated complex convolution, which combines complex convolution with dilated convolution. This structure can construct receptive fields at different scales and simultaneously model amplitude and phase information. This design effectively alleviates image distortion and noise amplification without introducing additional parameters. Figure 3As shown in (b), this branch employs three parallel dilated complex convolutions with dilation rates of 1, 2, and 4, respectively. This parallel structure, without changing the number of parameters, balances local details with global context, achieving efficient extraction of multi-scale frequency domain features. Each dilated complex convolution is sequentially connected to a Complex ReLU activation function and Complex Group Normalization (Complex GN) to enhance low-frequency feature response, improve the stability of brightness distribution, strengthen the expression of dark area details, and effectively suppress high-frequency noise interference. Finally, the concatenated multi-scale frequency domain features are transformed back to the spatial domain using IFFT. This process can be represented as:

[0059] Formula 4

[0060] Where W1, W2, and W3 are the features after processing by complex convolution branches with different hole rates, and their calculation formulas are as follows:

[0061] Formula 5

[0062] in, Represents global features. , and These represent dilated complex convolutions with dilation rates of 1, 2, and 4, respectively. This represents a convolution with a kernel size of 1×1.

[0063] Local features Y of the spatial branch output Spatial Global feature Y of frequency domain branch output Frequency After concatenation, the fused features are recalibrated using an efficient channel attention (ECA) mechanism, selecting the channels with the richest information for subsequent reconstruction. Finally, the recalibrated features are added to the features of the FreqEnhanceNet-enhanced image via residual connections to obtain the output features of the DDFM module. This achieves effective feature enhancement while maintaining gradient propagation stability. The process can be represented as: Formula 6, where, It is a FreqEnhanceNet-enhanced image. Represents convolution. For efficient channel attention, Concat represents a join operation.

[0064] In some embodiments, generating amplitude attention map features corresponding to the first branch image using the amplitude attention mechanism includes: performing color space transformation on the first branch image to obtain a transformed first branch image; decomposing the transformed first branch image to obtain an illuminance map; extracting multi-scale luminance features from the illuminance map to obtain luminance features; taking the amplitude spectrum of the luminance features to obtain the luminance distribution corresponding to the luminance features; and generating amplitude attention map features corresponding to the first branch image based on the luminance distribution.

[0065] In this embodiment, as Figure 4 As shown, the design of the Amplitude Attention Module (AAM) is inspired by the fact that the luminance component Y in the YCbCr color space concentrates the main luminance information of the image and has good noise robustness. AAM first extracts the Y component of the YCbCr color space from the first branch image. input The first branch image undergoes color space transformation to obtain the transformed first branch image. Based on Retinex theory, it is decomposed into an illuminance map to characterize the overall brightness distribution of the image. Subsequently, to efficiently extract multi-scale brightness features, multi-scale brightness feature extraction is performed on the illuminance map to obtain brightness features. The illuminance map is input into three parallel branches. Each branch employs depthwise separable convolutions (DSConv) at different scales, significantly reducing the number of parameters and computational complexity while extracting spatial features. Group normalization (GN) and ReLU activation functions are used to synergistically improve the model's stability, convergence efficiency, and feature representation capability. Finally, the output features of each branch are concatenated along the channel dimension to achieve the fusion of multi-scale contextual information, forming a comprehensive feature representation that is more robust to image illumination. This process can be represented as:

[0066] Formula 7 Formula 8

[0067] in, This represents a depthwise separable convolution operation. Normalization operation for representative groups Represents activation function operation. This represents the splicing operation along the channel dimension. G represents the illuminance map, which is the output brightness characteristic after connection.

[0068] right After performing an FFT, the amplitude spectrum is extracted to obtain the brightness distribution corresponding to the brightness features. This amplitude spectrum comprehensively reflects the brightness distribution characteristics of the image. Then, after processing with a 1×1 convolution, LeakyReLU, and Sigmoid activation function, amplitude attention map features are generated. This process can be represented as:

[0069] Formula Nine

[0070] in, To represent a 1×1 convolution operation, the Sigmoid activation function will... The value is constrained to the interval (0, 1). This represents an improved ReLU activation function. This indicates the amplitude operation.

[0071] In some embodiments, generating phase attention map features corresponding to the second branch image using the phase attention mechanism includes: performing a phase transformation on the second branch image to obtain a transformed second branch image; extracting multi-scale structural features from the transformed second branch image to obtain structural features; performing phase spectral extraction on the structural features to obtain structural contour features; and generating phase attention map features corresponding to the second branch image based on the structural contour features.

[0072] In this embodiment, the Phase Attention Module (PAM) module structure is as follows: Figure 5 As shown. The design of PAM is inspired by the principle of phase consistency, which effectively suppresses interference from illumination and grayscale changes while enhancing key structural information such as contours, edges, and textures. PAM first performs a phase transformation on the second branch image to obtain the transformed second branch image, i.e., extracts the phase consistency map from the second branch image. To obtain a spatial structure and contour representation with illumination invariance, the transformed second branch image is then subjected to multi-scale structural feature extraction to obtain structural features. The data is fed into three multi-directional Gabor convolution modules (MGCMs) dedicated to enhancing texture and contour features. MGCMs leverage the inherent advantages of two-dimensional Gabor filters in terms of orientation and scale awareness to further extract multi-scale, orientation-sensitive texture features. The calculation formula for two-dimensional Gabor filtering is as follows:

[0073] Formula 10

[0074] Formula 11 , Formula 12,

[0075] Wherein, λ, θ, φ, σ, and γ represent wavelength, direction, phase shift, standard deviation, and spatial aspect ratio, respectively; , ) are the image coordinates, ( , )for( , The result obtained by rotating in the θ direction, g() represents the Gabor filter response function, and i represents the complex exponential oscillation term.

[0076] MGCM structure as follows Figure 6 As shown, this module employs a parallel architecture, performing two-dimensional Gabor filtering in three directions: 0°, 60°, and 120°. The filtering response in each direction is sequentially processed through 3×3 convolution, 5×5 convolution, group normalization, and ReLU activation to achieve nonlinear transformation and multi-scale feature fusion. This module can simultaneously capture texture patterns and edge structures across multiple directions and scales, significantly enhancing its ability to express directional texture features. This process can be represented as:

[0077] Formula Thirteen

[0078] in, For a two-dimensional Gabor filter with direction θ, Represents a two-dimensional convolution operation. This represents a convolution operation with a kernel size of 5×5. This represents a convolution operation with a kernel size of 3×3. The final output feature map representing the θ-direction branch. It is a linear activation function. This is the group normalization function.

[0079] Output characteristics in the three directions of 0°, 60° and 120° , and By concatenating the components, the final output is obtained. The calculation formula is: Formula Fourteen

[0080] right An FFT is performed to extract the phase spectrum, capturing the structural contour features of the image. Subsequently, this phase spectrum undergoes a nonlinear transformation via 1×1 convolution and a LeakyReLU activation function, followed by normalization along the spatial dimension using a Softmax function to generate a phase attention map feature. This process can be represented as: Formula 15

[0081] in, Indicates to After performing FFT, the phase spectrum is obtained. Softmax represents the normalization function along the spatial dimension. Its output is constrained to a discrete probability distribution with values ​​in the interval (0,1) and a sum of 1. This is an improved ReLU activation function.

[0082] In some embodiments, determining the image brightness enhancement feature corresponding to the first branch image based on the initial amplitude spectrum feature and the amplitude attention map feature includes: multiplying the initial amplitude spectrum feature and the amplitude attention map feature element-wise to obtain a first weighted feature; and concatenating the first weighted feature with the initial amplitude spectrum feature using residual concatenation to obtain the image brightness enhancement feature.

[0083] In this embodiment, the amplitude attention map features are... As spatial weights, they are related to the initial amplitude spectrum features. Element-wise multiplication is performed to achieve amplitude modulation and adaptive selection and enhancement of key illumination features, improving the model's robustness under complex illumination conditions. Finally, the weighted amplitude features are connected via residuals. The fusion yields the enhanced amplitude characteristics. The calculation formula is Formula Sixteen

[0084] in, This indicates element-wise multiplication. This represents the weighted amplitude characteristics. This mechanism effectively avoids amplitude overflow in the Fourier space and can precisely control the amplitude enhancement intensity in different regions, thereby effectively increasing the brightness of underexposed areas while avoiding over-enhancement of normally exposed areas.

[0085] In some embodiments, determining the pixel structure enhancement features corresponding to the second branch image based on the initial phase features and the phase attention map features includes: multiplying the initial phase features and the phase attention map features element-wise to obtain a second weighted feature; and concatenating the second weighted feature with the initial phase features using residual concatenation to obtain the pixel structure enhancement features.

[0086] In this embodiment, the phase attention map features are... With initial phase characteristics Element-wise multiplication is performed to achieve phase modulation of key structural features in the frequency domain, enhancing the model's ability to represent contours and textures. Finally, the modulated phase features are connected via residuals. The phase features are fused together to obtain enhanced phase characteristics. The calculation formula is: Formula 17

[0087] in, This indicates element-wise multiplication. This represents the modulated phase characteristics. This design achieves optimized phase structure stability, enhancing phase consistency in key regions while effectively suppressing phase distortion in other regions.

[0088] In some embodiments, the loss function used to train the augmented model is determined by the following formula: ,in, Represents the loss function. This indicates the magnitude loss guided by illuminance. For illuminance adaptive weighting coefficients, This represents the frequency-domain weighted cosine phase loss. These are frequency domain adaptive weighting coefficients. For the histogram loss in the spatial domain, These are spatial adaptive weighting coefficients. In order to perceive loss, For sensing adaptive weighting coefficients.

[0089] In this embodiment, this application proposes a joint loss function to optimize model training by constraining the consistency between the enhanced image and the ground truth (GT) image in terms of frequency domain amplitude, phase distribution, spatial domain brightness, and semantic structure. Compared to existing technologies, the innovation of the loss function in this application lies in introducing the illumination-guided amplitude loss proposed in this application. ) and frequency-domain weighted cosine phase loss ( ), and incorporates the spatial histogram loss from existing technologies ( and perceived loss ( The total loss function is: Formula 18, where λ1 to λ4 are hyperparameters. Experimental analysis shows that the model performance is optimal when λ1=0.4, λ2=0.4, λ3=0.1, and λ4=0.1.

[0090] The following describes the amplitude loss and phase loss functions proposed in this application. (I) Illumination-guided amplitude loss: To achieve effective alignment of the enhanced image and the GT image in the amplitude spectrum, this application proposes an illumination-guided amplitude loss, the calculation formula of which is as follows: Formula 19, where, This represents the enhanced image. G represents the amplitude spectrum of the GT image; G(u,v) is the illuminance map adjusted in the frequency domain. The constant is used to avoid the denominator being zero, and N is the total number of pixels in the frequency domain feature map. The horizontal frequency coordinate is used. For the vertical frequency coordinates. This loss achieves amplitude distribution alignment through the following dual mechanisms: (1) Illumination adaptive weighting: A weight term (1-G(u,v)) is used, with a larger weight assigned to low-light areas (G(u,v)≈0) to increase brightness, and a smaller weight assigned to well-lit areas (G(u,v)≈1) to prevent overexposure and loss of detail. (2) Scale-invariant logarithmic difference: Logarithmic form It has scale invariance, making it more sensitive to relative changes in low-light areas, and effectively constrains the consistency of amplitude distribution between enhanced images and GT images.

[0091] (ii) Frequency-domain weighted cosine phase loss: To effectively preserve key structural features of the image and suppress artifact generation, this application proposes a frequency-domain weighted cosine phase loss to constrain the consistency of the enhanced image and the GT image in the phase spectrum. Its calculation formula is as follows: Formula 20, in which, This represents the enhanced image. Represents the phase spectrum of the GT image. The horizontal frequency coordinate is used. The vertical frequency coordinate is used. Cosine term. Used to assess phase consistency differences; the smaller the phase difference, the better. The lower the value. These are frequency-domain adaptive weights, used to assign higher weights to frequency-domain components that contribute more to the edge structure. Their calculation formula is: Formula 21, where, The Fourier transform representing the spatial gradient magnitude of the true image is used to quantify the importance of each frequency domain component to the edge structure; exponent For hyperparameters, when The model performs optimally at this time. The normalization factor R is the maximum radius, calculated as follows:

[0092] Formula 22.

[0093] In some embodiments, the training data for training the enhancement model is determined by the following method: acquiring multiple historical mine images, wherein the brightness of each historical mine image is less than or equal to a predetermined brightness; in response to determining that the number of the multiple historical mine images is less than a predetermined number, acquiring multiple normal images with brightness greater than the predetermined brightness; for each normal image, segmenting the normal image to obtain multiple superpixel regions; for each superpixel region corresponding to each normal image, randomly generating an attenuation coefficient for the superpixel region, and performing brightness attenuation on the superpixel region using Gaussian smoothing based on the attenuation coefficient; in response to determining that the brightness of the superpixel region after brightness attenuation is less than or equal to the predetermined brightness, using the superpixel region after brightness attenuation as a new image; in response to determining that the total number of the multiple historical mine images and all new images is greater than or equal to the predetermined number, using the multiple historical mine images and all new images together as training data for the enhancement model.

[0094] In this embodiment, the research on underground coal mine image enhancement is limited by the scarcity of real paired data. To this end, this application proposes a data simulation method to construct a coal mine underground low-light image dataset TG-ULLID containing 2900 pairs of images. Each pair of images consists of a normal image and its corresponding synthetic low-light image. The sources of the normal images include the following three parts: (1) 1200 high-quality images were selected from the DsLMF+ dataset based on information entropy, average gradient and NIQE index. (2) 1200 high-quality images were selected from the CUMT_belt underground conveyor belt foreign object detection dataset using the same evaluation criteria. (3) 500 coal and rock sample images were obtained through laboratory collection.

[0095] When the number of historical mine images is less than a predetermined number, more training images need to be added to ensure the accuracy of model training. To simulate the non-uniform lighting environment in coal mines, an image synthesis method based on Simple Linear Iterative Clustering (SLIC) is adopted. This method first segments the normal image into multiple superpixel regions, then randomly generates a brightness attenuation coefficient for each region, and uses Gaussian smoothing to achieve a natural lighting transition between different regions. The calculation formula is as follows: Formula 23, where, For the generated low-light image, Standard deviation A Gaussian convolution kernel with a kernel size of 5×5. This represents the convolution operation. For the original image, This indicates element-wise multiplication, where M is the initial brightness attenuation mask generated based on the brightness attenuation coefficient of the superpixel region. Gaussian smoothing allows for a natural transition in brightness attenuation between different regions, effectively simulating non-uniform lighting conditions downhole.

[0096] The training data for the enhanced model employs a strictly randomized partitioning principle, ensuring that the image content across different sets is independent. For example, this results in a training set of 2320 image pairs, a validation set of 290 image pairs, and a test set of 290 image pairs. Furthermore, this application additionally collects 50 pairs of real underground coal mine images to construct an independent benchmark set for impartially evaluating model performance and verifying its generalization ability in the real-world environment of underground coal mines.

[0097] Experimental images were uniformly scaled to 608×608 pixels. Training employed data augmentation techniques such as random cropping, flipping, and multi-scale augmentation, with a batch size of 8. The SGD optimizer (initial learning rate 0.001, weight decay 0.0005) was used, and training was performed for 30 epochs on an RTX 3090 GPU within the PyTorch framework.

[0098] In another embodiment provided in this application, a comprehensive comparative analysis was conducted on the self-made underground coal mine dataset TG-ULLID, comparing the method of this application with a variety of advanced methods (including Retinexformer, Retinexmamba, FIPNet, FourLLIE, FENet and CF-UFormer, etc.) from 2023 to 2025 from both subjective visual perception and objective evaluation indicators.

[0099] Subjective visual contrast effect, such as Figure 7 As shown, UHDFour suffers from insufficient color saturation, resulting in a grayish conveyor belt and failing to fully restore the surface texture of the hydraulic support. FourLLIE exhibits significant color shift, manifested as oversaturation of the safety helmet color and the introduction of unnatural shadows on the tunnel walls. While Retinexformer and Retinexmamba can improve overall brightness, their saturation and contrast remain insufficient, leading to darker brightness in key underground areas. FIPNet suffers from exposure imbalance, exhibiting both local overexposure and underexposure; FENet, while improving brightness, suffers from insufficient overall illumination recovery and low detail recognition in dark areas. In contrast, the method in this application achieves adaptive brightness adjustment and local contrast preservation in the underground environment through frequency domain co-optimization; it effectively suppresses dust and noise interference, significantly improving the fidelity of conveyor belt edges and tunnel textures. Subjective enhancement results show that the enhancement results of the method in this application are visually closest to the real scene, demonstrating significant advantages in color reproduction, detail clarity, and illumination balance.

[0100] In terms of objective evaluation, Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Natural Image Quality Evaluator (NIQE) were used as image quality evaluation metrics. The computational complexity and real-time performance of the methods were measured using model parameter count, number of floating-point operations (FLOPs), and single-image testing time. Specific comparison results are shown in Table 1.

[0101] Table 1. Comparison of objective metrics for the TG-ULLID dataset (bold indicates the best result, and underline indicates the second best result).

[0102]

[0103] As shown in Table 1, the proposed method achieves a PSNR of 25.85 dB, which is 1.03 dB higher than the second-best method, UHDFour; it also achieves an SSIM score of 0.859, ranking first; and it achieves an NIQE score of 3.59, second only to the best method, FENet. Furthermore, the proposed method demonstrates excellent performance in terms of model complexity, maintaining high performance while also possessing good computational efficiency: the model has only 0.18 M parameters, 15.25 G FLOPs, and a single image processing time of 0.103 seconds, with overall complexity superior to most mainstream comparison methods. Experimental results show that the proposed method exhibits excellent image enhancement performance in low-light scenes in coal mines, effectively suppressing noise while achieving natural visual enhancement, achieving a good balance between improving image quality and maintaining visual realism.

[0104] In another embodiment provided in this application, comparative experiments are conducted on the LOL-v2 public dataset to verify the generalization performance of the method. Subjective visual comparison, for example... Figure 8 As shown. By Figure 8It can be seen that existing methods have significant limitations: UHDFour exhibits edge distortion, FourLLIE suffers from insufficient saturation and color deviation, Retinexformer and Retinexmamba show overall insufficient contrast and localized darkness, CF-UFormer exhibits over-enhancement and noise amplification, FIPNet improves color fidelity but introduces low-frequency noise when enhancing dark areas, and FENet shows uneven brightness recovery. In contrast, the APSFNet proposed in this paper effectively suppresses noise while significantly improving edge preservation and texture detail restoration capabilities, demonstrating superior performance in color fidelity, detail reproduction quality, and brightness distribution uniformity. Quantitative analysis results are shown in Table 2. The proposed method achieves a PSNR of 23.95 dB on the LOL-v2-Real subset, a 0.86 dB improvement over the second-best method, UHDFour, and also ranks first in NIQE. On the LOL-v2-Synthetic subset, it achieves a PSNR of 26.05 dB, an SSIM of 0.941, and an NIQE of 3.61, all ranking first in all three metrics. Experimental results demonstrate that the proposed method exhibits excellent generalization ability on public datasets and achieves a good balance between noise suppression and natural enhancement.

[0105] Table 2 Comparison of objective results for low-light images in the LOL-v2 dataset (optimal results are marked in bold, and suboptimal results are marked in underline).

[0106]

[0107] In another embodiment provided in this application, to verify the effect of image enhancement methods on target detection performance in low-light environments in coal mines, image enhancement methods are used as preprocessing methods and then input into the YOLOv12 model to detect underground workers and conveyor belt foreign objects in the TG-ULLID dataset. Simultaneously, to increase the comprehensiveness of the comparison, MPE-DETR based on Transformer is added as a comparison method. The performance evaluation metrics of each detection method are shown in Table 3, mainly including precision (P), recall (R), frames per second (FPS), and the comprehensive evaluation metric F-score. The F-score is defined as the harmonic mean of precision and recall, aiming to comprehensively evaluate the overall performance of the model.

[0108] Table 3. Target detection results of different detection methods on the TG-ULLID dataset (optimal results are marked in bold, and suboptimal results are marked in underline).

[0109]

[0110] Compared to other methods, the image enhanced by the method described in this application achieves the most accurate detection results due to clearer detail and better suppression of background noise. Table 3 further verifies this advantage: on the TG-ULLID dataset, although various enhancement methods bring certain performance improvements, the method described in this application performs best in several key metrics. Specifically, the method described in this application ranks first in both P and F scores, reaching 87.8% and 89.1% respectively; the R score is 90.5%, ranking second among all methods, second only to the FENet method's 90.8%. In addition, while maintaining high accuracy, the method described in this application has a frame rate of 87.15, demonstrating excellent real-time processing capabilities and proving its practicality and robustness in low-light environments in coal mines.

[0111] In another embodiment provided in this application, to comprehensively evaluate the performance of the method, a coal mine safety helmet detection experiment was conducted on the public dataset DsLMF+helmet. A "first augmentation, then detection" approach was adopted to verify the generalization ability of the method in different underground scenarios. By analyzing the precision-recall curves and subjective detection results, the actual improvement effect of each augmentation method on downstream detection tasks was evaluated. Experimental results are as follows: Figure 9 As shown in the precision-recall curves, the method in this application achieves the largest area under the curve, and its overall detection performance is superior to all comparative methods.

[0112] In another embodiment provided in this application, ablation experiments were conducted to evaluate the contribution of each module and loss function to the model performance. The ablation experiment results on the TG-ULLID and LOL-v2 datasets are shown in Tables 4 and 5, respectively. It can be seen that the complete APSFNet achieves the best performance in all test scenarios. Specifically, on the TG-ULLID dataset, compared with Model 1 (without PAM), Model 2 (without AAM), and Model 3 (without DDFM), the complete model improves PSNR by 1.03dB, 1.20dB, and 1.50dB, respectively; SSIM by 0.027, 0.018, and 0.021, respectively; and NIQE by 0.43, 0.46, and 0.61, respectively. On the LOL-v2-Real dataset, the full model improved PSNR by 1.03 dB, 1.10 dB, and 1.50 dB, respectively; SSIM improved by 0.039, 0.024, and 0.029; and NIQE decreased by 0.37, 0.39, and 0.54. A consistent trend was observed on the LOL-v2-Synthetic dataset, with PSNR improvements of 0.70 dB, 0.90 dB, and 1.30 dB, SSIM improvements of 0.024, 0.012, and 0.021, and NIQE decreases of 0.17, 0.19, and 0.27. These results indicate that DDFM makes the most significant contribution to image quality improvement, PAM plays a crucial role in structure preservation, and AAM effectively improves the overall brightness uniformity of the image. The three work synergistically to ensure the superior performance of APSFNet on different datasets.

[0113] Table 4 Comparison of ablation experiments on the TG-ULLID core module (optimal results are marked in bold).

[0114]

[0115] Table 5 Comparison of ablation experiments on LOL-v2 core modules (optimal results are marked in bold).

[0116]

[0117] The results of the loss function ablation experiments are shown in Tables 6 and 7. On the TG-ULLID and LOL-v2 datasets, the complete model significantly outperformed the versions with each loss function removed, with LAmp and LPha making the most crucial contributions: removing LAmp resulted in a decrease in PSNR of 1.60 to 1.90 dB, a decrease in SSIM of 0.043 to 0.069, and an increase in NIQE of 0.09 to 0.19; removing LPha resulted in a decrease in PSNR of 1.27 to 1.77 dB, a decrease in SSIM of 0.064 to 0.115, and an increase in NIQE of 0.04 to 0.14. Furthermore, the removal of LPer and LHist also caused a decrease in model performance, further validating the necessity of all loss functions in improving model performance.

[0118] Table 6 Comparison of ablation experiments for different loss functions in TG-ULLID (optimal results are marked in bold).

[0119]

[0120] Table 7 Comparison of ablation experiments for different loss functions in LOL-v2 (optimal results are marked in bold).

[0121]

[0122] In another embodiment provided in this application, focusing on the key issue of low-light image enhancement in coal mines, this application proposes an image enhancement method, APSFNet, which integrates amplitude and phase features with spatial and frequency domain features. Experimental results on the self-built TG-ULLID and the publicly available LOL-v2 datasets show that the proposed method outperforms existing mainstream methods in key evaluation metrics such as PSNR, SSIM, and NIQE, especially in terms of brightness enhancement, detail restoration, and noise suppression. As a preprocessing module, APSFNet significantly improves the performance of underground target detection tasks, verifying its practical application value in intelligent mine construction.

[0123] The core contributions of this application mainly include: (1) Establishing an amplitude and phase co-enhancement mechanism, and achieving an effective balance between image brightness enhancement and structural feature preservation by constructing a dual-path attention module. (2) Designing a spatial and frequency domain dual-domain fusion architecture, and constructing a complementary mechanism between spatial local details and frequency global context, which significantly improves the quality of detail recovery and suppresses noise interference. (3) Proposing a joint loss function, as well as illumination-guided amplitude loss and frequency-domain weighted cosine phase loss, which effectively ensures the consistency of enhancement results in terms of visual quality and structural integrity. (4) Constructing the coal mine underground image enhancement dataset TG-ULLID.

[0124] It should be noted that the method in this embodiment can be executed by a single device, such as a computer or server. The method can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method in this embodiment, and the multiple devices will interact with each other to complete the method described.

[0125] It should be noted that the above description describes some embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0126] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides a mine image enhancement device.

[0127] refer to Figure 10 The mine image enhancement device includes:

[0128] The decomposition module 10 is configured to acquire a mine image and input the mine image into a pre-trained enhancement model, and perform transformation decomposition on the mine image through the enhancement model to obtain a first branch image and a second branch image.

[0129] The first generation module 20 is configured to extract the amplitude spectrum of the first branch image to obtain initial amplitude spectrum features, and to generate amplitude attention map features corresponding to the first branch image using an amplitude attention mechanism.

[0130] The first determining module 30 is configured to determine the image brightness enhancement features corresponding to the first branch image based on the initial amplitude spectrum features and the amplitude attention map features.

[0131] The second generation module 40 is configured to perform phase extraction on the second branch image to obtain initial phase features, and to generate phase attention map features corresponding to the second branch image using a phase attention mechanism.

[0132] The second determining module 50 is configured to determine the pixel structure enhancement features corresponding to the second branch image based on the initial phase features and the phase attention map features.

[0133] The third determining module 60 is configured to determine the enhanced mine image based on the image brightness enhancement features and the pixel structure enhancement features, and output the enhanced mine image through the enhancement model.

[0134] Using the aforementioned apparatus, a mine image is acquired and input into a pre-trained enhancement model. The enhancement model performs transform decomposition on the mine image to obtain a first branch image and a second branch image. Amplitude spectrum extraction is performed on the first branch image to obtain initial amplitude spectrum features. An amplitude attention map feature corresponding to the first branch image is generated using an amplitude attention mechanism. Based on the initial amplitude spectrum features and the amplitude attention map feature, image brightness enhancement features corresponding to the first branch image are determined. Phase extraction is performed on the second branch image to obtain initial phase features. A phase attention map feature corresponding to the second branch image is generated using a phase attention mechanism. Based on the initial phase features and the phase attention map feature, pixel structure enhancement features corresponding to the second branch image are determined. Based on the image brightness enhancement features and the pixel structure enhancement features, an enhanced mine image is determined, and the enhanced mine image is output through the enhancement model.

[0135] In some embodiments, the third determining module 60 is further configured to transform and concatenate the image brightness enhancement features and the pixel structure enhancement features to obtain concatenated features; use the encoder in the enhancement model to perform multi-level spatial feature extraction on the concatenated features to obtain multiple enhanced concatenated features; use the cascaded multi-scale dynamic serpentine convolution in the enhancement model to extract the local features corresponding to each enhanced concatenated feature; use the multi-hole dilated complex convolution in the enhancement model to model the multiple enhanced concatenated features to obtain global features; and use the decoder in the enhancement model to perform image reconstruction on the global features and the local features corresponding to each enhanced concatenated feature to obtain the enhanced mine image.

[0136] In some embodiments, the first generation module 20 is further configured to perform color space conversion on the first branch image to obtain a converted first branch image; decompose the converted first branch image to obtain an illuminance map; extract multi-scale luminance features from the illuminance map to obtain luminance features; extract the amplitude spectrum from the luminance features to obtain the luminance distribution corresponding to the luminance features; and generate amplitude attention map features corresponding to the first branch image based on the luminance distribution.

[0137] In some embodiments, the second generation module 40 is further configured to perform phase transformation on the second branch image to obtain a transformed second branch image; extract multi-scale structural features from the transformed second branch image to obtain structural features; perform phase spectrum extraction on the structural features to obtain structural contour features; and generate phase attention map features corresponding to the second branch image based on the structural contour features.

[0138] In some embodiments, the first determining module 30 is further configured to perform element-wise multiplication of the initial amplitude spectrum feature and the amplitude attention map feature to obtain a first weighted feature; and to perform residual concatenation of the first weighted feature and the initial amplitude spectrum feature to obtain the image brightness enhancement feature.

[0139] In some embodiments, the second determining module 50 is further configured to perform element-wise multiplication of the initial phase feature and the phase attention map feature to obtain a second weighted feature; and to perform residual concatenation of the second weighted feature and the initial phase feature to obtain the pixel structure enhancement feature.

[0140] In some embodiments, a first training module is further included, the first training module being configured to determine the loss function for training the augmented model by the following formula: ,in, Represents the loss function. This indicates the magnitude loss guided by illuminance. For illuminance adaptive weighting coefficients, This represents the frequency-domain weighted cosine phase loss. These are frequency domain adaptive weighting coefficients. For the histogram loss in the spatial domain, These are spatial adaptive weighting coefficients. In order to perceive loss, For sensing adaptive weighting coefficients.

[0141] In some embodiments, a second training module is further included, the second training module being configured to determine the training data for training the enhancement model by the following method: acquiring a plurality of historical mine images, wherein the brightness of each historical mine image is less than or equal to a predetermined brightness; in response to determining that the number of the plurality of historical mine images is less than a predetermined number, acquiring a plurality of normal images with brightness greater than the predetermined brightness; for each normal image, segmenting the normal image to obtain a plurality of superpixel regions; for each superpixel region corresponding to each normal image, randomly generating an attenuation coefficient for the superpixel region, and performing brightness attenuation on the superpixel region using Gaussian smoothing based on the attenuation coefficient; in response to determining that the brightness of the superpixel region after brightness attenuation is less than or equal to the predetermined brightness, using the superpixel region after brightness attenuation as a new image; in response to determining that the total number of the plurality of historical mine images and all new images is greater than or equal to the predetermined number, using the plurality of historical mine images and all new images together as training data for the enhancement model.

[0142] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware.

[0143] The apparatus of the above embodiments is used to implement the corresponding mine image enhancement method in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0144] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the mine image enhancement method as described in any of the above embodiments.

[0145] Figure 11 This embodiment illustrates a more specific hardware structure of an electronic device. The device may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0146] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0147] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0148] The input / output interface 1030 is used to connect input / output modules to realize information input and output. The input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0149] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0150] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0151] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0152] The electronic devices described above are used to implement the corresponding mine image enhancement methods in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0153] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the mine image enhancement method as described in any of the above embodiments.

[0154] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0155] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the mine image enhancement method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0156] Based on the same concept, corresponding to the methods of any of the above embodiments, this application also provides a computer program product, including computer program instructions, which, when run on a computer, cause the computer to execute the mine image enhancement method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0157] It should be noted that the embodiments of this application can also be further described in the following ways:

[0158] It is understood that before using the technical solutions of the various embodiments in this disclosure, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.

[0159] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose, based on the prompt message, whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media performing the operations of this disclosed technical solution.

[0160] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0161] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0162] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0163] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0164] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0165] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. A method for enhancing mine images, characterized in that, include: A mine image is acquired and input into a pre-trained enhancement model. The enhancement model then performs a transform decomposition on the mine image to obtain a first branch image and a second branch image. Amplitude spectrum extraction is performed on the first branch image to obtain initial amplitude spectrum features, and amplitude attention map features corresponding to the first branch image are generated using an amplitude attention mechanism. Based on the initial amplitude spectrum features and the amplitude attention map features, determine the image brightness enhancement features corresponding to the first branch image; Phase extraction is performed on the second branch image to obtain initial phase features, and phase attention map features corresponding to the second branch image are generated using a phase attention mechanism; Based on the initial phase features and the phase attention map features, determine the pixel structure enhancement features corresponding to the second branch image; Based on the image brightness enhancement features and the pixel structure enhancement features, an enhanced mine image is determined, and the enhanced mine image is output through the enhancement model; The step of determining the enhanced mine image based on the image brightness enhancement features and the pixel structure enhancement features includes: transforming and stitching the image brightness enhancement features and the pixel structure enhancement features to obtain stitched features; using the encoder in the enhancement model to perform multi-level spatial feature extraction on the stitched features to obtain multiple enhanced stitched features; using the cascaded multi-scale dynamic serpentine convolution in the enhancement model to extract the local features corresponding to each enhanced stitched feature; using the multi-void complex convolution with multiple void ratios in the enhancement model to model the multiple enhanced stitched features to obtain global features; and using the decoder in the enhancement model to perform image reconstruction on the global features and the local features corresponding to each enhanced stitched feature to obtain the enhanced mine image.

2. The method according to claim 1, characterized in that, The step of generating amplitude attention map features corresponding to the first branch image using the amplitude attention mechanism includes: The first branch image is color space converted to obtain the converted first branch image; The transformed first branch image is decomposed to obtain an illumination map; Multi-scale luminance feature extraction is performed on the illuminance map to obtain luminance features; The amplitude spectrum of the brightness feature is taken to obtain the brightness distribution corresponding to the brightness feature; Based on the brightness distribution, amplitude attention map features corresponding to the first branch image are generated.

3. The method according to claim 1, characterized in that, The step of generating phase attention map features corresponding to the second branch image using the phase attention mechanism includes: Perform a phase transformation on the second branch image to obtain the transformed second branch image; Multi-scale structural feature extraction is performed on the transformed second branch image to obtain structural features; Phase spectrum extraction is performed on the structural features to obtain the structural contour features; Based on the structural contour features, the phase attention map features corresponding to the second branch image are generated.

4. The method according to claim 1, characterized in that, The step of determining the image brightness enhancement features corresponding to the first branch image based on the initial amplitude spectrum features and the amplitude attention map features includes: The initial amplitude spectrum feature and the amplitude attention map feature are multiplied element-wise to obtain the first weighted feature; The image brightness enhancement feature is obtained by concatenating the first weighted feature with the initial amplitude spectrum feature using residual concatenation.

5. The method according to claim 1, characterized in that, The step of determining the pixel structure enhancement features corresponding to the second branch image based on the initial phase features and the phase attention map features includes: The initial phase feature and the phase attention map feature are multiplied element-wise to obtain the second weighted feature; The second weighted feature is concatenated with the initial phase feature using residual concatenation to obtain the pixel structure enhancement feature.

6. The method according to claim 1, characterized in that, The loss function used to train the augmented model is determined by the following formula: ， in, Represents the loss function. This indicates the magnitude loss guided by illuminance. For illuminance adaptive weighting coefficients, This represents the frequency-domain weighted cosine phase loss. These are frequency domain adaptive weighting coefficients. For the histogram loss in the spatial domain, These are spatial adaptive weighting coefficients. In order to perceive loss, For sensing adaptive weighting coefficients.

7. The method according to claim 1, characterized in that, The training data used to train the augmented model is determined by the following method: Acquire multiple historical mine images, wherein the brightness of each historical mine image is less than or equal to a predetermined brightness; In response to determining that the number of the plurality of historical mine images is less than a predetermined number, a plurality of normal images with a brightness greater than the predetermined brightness are acquired; For each normal image, the normal image is segmented to obtain multiple superpixel regions; For each superpixel region corresponding to each normal image, a decay coefficient for the superpixel region is randomly generated, and the brightness of the superpixel region is attenuated using Gaussian smoothing based on the decay coefficient. In response to determining that the brightness of the superpixel region after brightness attenuation is less than or equal to the predetermined brightness, the superpixel region after brightness attenuation is used as a new image. In response to determining that the total number of the plurality of historical mine images and all newly added images is greater than or equal to the predetermined number, the plurality of historical mine images and all newly added images are used together as training data for the enhancement model.

8. A device for enhancing mine images, characterized in that, include: The decomposition module is configured to acquire a mine image and input the mine image into a pre-trained enhancement model, and perform transformation decomposition on the mine image through the enhancement model to obtain a first branch image and a second branch image; The first generation module is configured to extract the amplitude spectrum of the first branch image to obtain initial amplitude spectrum features, and to generate amplitude attention map features corresponding to the first branch image using an amplitude attention mechanism. The first determining module is configured to determine the image brightness enhancement features corresponding to the first branch image based on the initial amplitude spectrum features and the amplitude attention map features; The second generation module is configured to perform phase extraction on the second branch image to obtain initial phase features, and to generate phase attention map features corresponding to the second branch image using a phase attention mechanism. The second determining module is configured to determine the pixel structure enhancement features corresponding to the second branch image based on the initial phase features and the phase attention map features; The third determining module is configured to determine the enhanced mine image based on the image brightness enhancement features and the pixel structure enhancement features, and output the enhanced mine image through the enhancement model; The third determining module is further configured to transform and splice the image brightness enhancement feature and the pixel structure enhancement feature to obtain spliced ​​features; Using the encoder in the enhancement model, multi-level spatial feature extraction is performed on the splicing features to obtain multiple enhanced splicing features; using the cascaded multi-scale dynamic serpentine convolution in the enhancement model, the local features corresponding to each enhanced splicing feature are extracted; By utilizing the multi-hole complex convolution with multiple void ratios in the enhancement model, multiple enhanced splicing features are modeled to obtain global features; Using the decoder in the enhancement model, the global features and the local features corresponding to each enhanced stitched feature are used to restore the image, thereby obtaining the enhanced mine image.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

Images