DeepPatchNet: A FFT-CNN-Based Fusion Approach for Resilient Phase Correction in Underwater Image Reconstruction

This paper introduces a groundbreaking methodology for phase correction in underwater images, utilizing a novel Convolutional Neural Network (CNN) model developed through a patch-based approach. The effectiveness of the proposed method is rigorously evaluated using Peak Signal-to-Noise Ratio (PSNR) and Mean Squared Error (MSE) analyses. Results from the reconstruction of 10 images showcase a significant improvement compared to conventional Fast Fourier Transform (FFT)-based methods, achieving a maximum PSNR of 35.6 dB and an impressive average PSNR of 34 dB. MSE analysis indicates optimal performance, particularly under a 10% attack, with a minimum value of 0.28. However, under 20% and 30% attack scenarios, a slightly diminished PSNR is observed compared to the more favorable 10% scenario. The proposed CNN-based patch approach not only surpasses traditional FFT methods in performance but also exhibits resilience to varying degrees of attack, marking it as a promising advancement in underwater image phase correction.