<p>Due to the limited lighting conditions in the mine environment, the captured images often exhibit blurriness, dimness, and lack of detail. To address these issues, a mine image super-resolution reconstruction method based on brightness attention and dual-stream self-attention is proposed. First, a brightness attention map is introduced to guide the network in enhancing the brightness of dark areas while ensuring that bright areas maintain appropriate brightness to prevent overexposure. Second, a multi-scale hierarchical fusion module is designed to extract image features at different scales, integrating low-level features into high-level features in a top-down manner to further improve the detail representation of super-resolution images. Then, a spatial-channel self-attention parallel interaction module is adopted to enhance the complementarity between different scales and levels of features, allowing the network to focus more on the detailed information in the low-level features and the semantic information in the high-level features. Additionally, grouped densely connected residual blocks are designed to enhance the model’s ability to extract and process rich feature information. To address the insufficient consideration of feature complementarity and continuity in the boundary regions of image patches in existing mine image super-resolution algorithms, which leads to poor detail representation in reconstructed images, a arbitrary size super-resolution based on shallow feature Gaussian aggregation algorithms is proposed. This method partitions shallow features, reducing computational cost while effectively solving the issues of artifacts and discontinuities in the boundary regions of image patches. Experimental results demonstrate that, on the public mine image dataset, the proposed method achieves the best quantitative evaluation metrics in <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\times $</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation>2, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\times $</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation>3, and <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$\times $</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation>4 super-resolution reconstruction tasks compared to other super-resolution algorithms. In the qualitative analysis, the super-resolved images generated by the proposed method exhibit superior visual quality, preserving image details and color representation to the greatest extent, with significant brightness enhancement in the dark areas.</p>

Super Resolution Reconstruction of Mine Images Based on Brightness Attention and Dual-Stream Self-Attention

期刊论文
发表日期：

2026年3月22日

查看全文

Xiaolei Zheng,
Weizhong Chu

摘要

Due to the limited lighting conditions in the mine environment, the captured images often exhibit blurriness, dimness, and lack of detail. To address these issues, a mine image super-resolution reconstruction method based on brightness attention and dual-stream self-attention is proposed. First, a brightness attention map is introduced to guide the network in enhancing the brightness of dark areas while ensuring that bright areas maintain appropriate brightness to prevent overexposure. Second, a multi-scale hierarchical fusion module is designed to extract image features at different scales, integrating low-level features into high-level features in a top-down manner to further improve the detail representation of super-resolution images. Then, a spatial-channel self-attention parallel interaction module is adopted to enhance the complementarity between different scales and levels of features, allowing the network to focus more on the detailed information in the low-level features and the semantic information in the high-level features. Additionally, grouped densely connected residual blocks are designed to enhance the model’s ability to extract and process rich feature information. To address the insufficient consideration of feature complementarity and continuity in the boundary regions of image patches in existing mine image super-resolution algorithms, which leads to poor detail representation in reconstructed images, a arbitrary size super-resolution based on shallow feature Gaussian aggregation algorithms is proposed. This method partitions shallow features, reducing computational cost while effectively solving the issues of artifacts and discontinuities in the boundary regions of image patches. Experimental results demonstrate that, on the public mine image dataset, the proposed method achieves the best quantitative evaluation metrics in $\times $ $\times$ 2, $\times $ $\times$ 3, and $\times $ $\times$ 4 super-resolution reconstruction tasks compared to other super-resolution algorithms. In the qualitative analysis, the super-resolved images generated by the proposed method exhibit superior visual quality, preserving image details and color representation to the greatest extent, with significant brightness enhancement in the dark areas.