Year: 2026
Type: preprint
Abstract: In railway fastener monitoring, poor lighting leads to issues like shadow distortion, overexposure, and loss of texture details, affecting system reliability and analysis accuracy. To address these, we propose the LSM-KinD model, which integrates the KinD architecture with SENet attention and multi-scale residual connections. The model uses an image decomposition network to separate reflection and illumination components, enhancing accuracy with multi-scale feature extraction and attention mechanisms. The illumination adjustment network applies residual blocks and channel attention to precisely control illumination, improving image quality. The image restoration network, based on U-Net and channel attention, optimizes the enhancement results. Experimental results show that LSM-KinD outperforms other algorithms, achieving PSNR and SSIM scores of 23.921 and 0.901 on the RAILF dataset. The model also improves detection accuracy, reaching a [email protected] of 84.9% in low-light environments, demonstrating its broad application potential in railway safety and system maintenance.
Source: Research Square
Language: English
Topic: Image Enhancement Techniques
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Award: G3285737361
OpenAlex ID: W7234293016
