An Unsupervised Image Enhancement Framework for Multiple Fault Detection of Insulators

To address the problem of low detection accuracy caused by uneven brightness distribution in transmission line inspection images under complex lighting conditions, this paper proposes an unsupervised image enhancement method that integrates grayscale feature guidance and luminance consistency loss constraint. First, a U-shaped generator combining a bottleneck structure with large receptive field depthwise separable convolutions is designed to efficiently extract multi-scale features. Second, a grayscale feature-guided image generation module is incorporated into the generator, using grayscale information to adaptively enhance local low-light regions and effectively suppress overexposed regions. Meanwhile, to accommodate the characteristics of unpaired data training, a luminance consistency loss is introduced. By constraining the global luminance distribution consistency between the generated image and the reference image, the overall brightness balance of the generated image is improved. Finally, a multi-level discriminator structure is constructed to enhance the model's ability to distinguish global and local luminance in the generated images. Experimental results show that the proposed method significantly improves image quality (PSNR increased from 7.73 to 18.41, SSIM increased from 0.43 to 0.85). Furthermore, the enhanced images lead to improvements in defect detection accuracy.