LM-CycleGAN: Improving Underwater Image Quality Through Learned Perceptual Image Patch Similarity and Multi-Scale...

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Underwater images suffer from degradation like noise, blur, and color distortion due to light scattering and particles.
• Existing image enhancement methods struggle with complex underwater environments.

Purpose of the Study:

• To propose an improved Cycle-consistent Generative Adversarial Network (CycleGAN) for underwater image color correction and detail enhancement.
• To address limitations of traditional CycleGAN in preserving structural information and avoiding artifacts.

Main Methods:

• Developed the LPIPS-MAFA CycleGAN (LM-CycleGAN) model.
• Integrated a Multi-scale Adaptive Fusion Attention (MAFA) mechanism for detail perception.
• Incorporated Learned Perceptual Image Patch Similarity (LPIPS) into the loss function for structural focus.

Main Results:

• LM-CycleGAN demonstrated significant improvements in SSIM, PSNR, AG, UCIQE, and UIQM on UIEB and EUVP datasets.
• The model achieved superior color correction and fidelity compared to traditional CycleGAN.
• Successfully avoided red checkerboard artifacts and blurred edge details.

Conclusions:

• LM-CycleGAN offers a robust solution for underwater image enhancement.
• The proposed model effectively improves both image quality and visual fidelity in challenging underwater conditions.