Underwater Image Enhancement using Convolution Denoising Network and Blind Convolution

Abstract

Underwater Image Enhancement (UWIE) is essential for improving the quality of Underwater Images (UWIs). However, recent UWIE methods face challenges due to low lighting conditions, contrast issues, color distortion, lower visibility, stability and buoyancy, pressure and temperature, and white balancing problems. Traditional techniques cannot capture the fine changes in UWI texture and cannot learn complex patterns. This study presents a UWIE Network (UWIE-Net) based on a parallel combination of a denoising Deep Convolution Neural Network (DCNN) and blind convolution to improve the overall visual quality of UWIs. The DCNN is used to depict the UWI complex pattern features and focuses on enhancing the image's contrast, color, and texture. Blind convolution is employed in parallel to minimize noise and irregularities in the image texture. Finally, the images obtained at the two parallel layers are fused using wavelet fusion to preserve the edge and texture information of the final enhanced UWI. The effectiveness of UWIE-Net was evaluated on the Underwater Image Enhancement Benchmark Dataset (UIEB), achieving MSE of 23.5, PSNR of 34.42, AG of 13.56, PCQI of 1.23, and UCIQE of 0.83. The UWIE-Net shows notable improvement in the overall visual and structural quality of UWIs compared to existing state-of-the-art methods.