An Unsupervised Image Enhancement Framework for Multiple Fault Detection of Insulators

Area of Science:

• Computer Vision
• Image Processing
• Electrical Engineering

Background:

• Transmission line inspection requires high-quality images.
• Complex lighting conditions cause uneven brightness, reducing detection accuracy.
• Existing methods struggle with unsupervised enhancement under varied lighting.

Purpose of the Study:

• To develop an unsupervised image enhancement method for transmission line inspection.
• To improve detection accuracy under complex lighting conditions.
• To address issues of uneven brightness distribution in inspection images.

Main Methods:

• Proposed an unsupervised method integrating grayscale feature guidance and luminance consistency loss.
• Designed a U-shaped generator with depthwise separable convolutions for multi-scale feature extraction.
• Incorporated a grayscale feature-guided module and a luminance consistency loss for adaptive enhancement and brightness balance.
• Utilized a multi-level discriminator for enhanced global and local luminance distinction.

Main Results:

• Significantly improved image quality metrics: Peak Signal-to-Noise Ratio (PSNR) increased from 7.73 to 18.41.
• Structural Similarity Index (SSIM) improved from 0.43 to 0.85.
• Enhanced images led to improved defect detection accuracy in transmission line inspections.

Conclusions:

• The proposed unsupervised method effectively enhances transmission line inspection images under complex lighting.
• Grayscale guidance and luminance consistency loss are crucial for adaptive enhancement and brightness balance.
• The method improves overall image quality and subsequent defect detection accuracy.