Enhancing Underwater Image Quality Using Multi-Scale Retinex for Improved Object Detection with YOLOv5 and Self-Supervised Learning with SimCLR

Object detection for underwater exploration is a critical process within the ambiance of comfort. However, it has various unique challenges such as identifying densely packed, small targets, dealing with noisy underwater images, scattering of light, and light absorption. In this paper, we propose a novel approach to detect and classify objects underwater. The underwater images are enhanced using Multi-Scale Retinex (MSR). SimCLR based modified YOLOv5 model with pretext tasked Resnet50 instead of Darknet50 is used as backbone. SimCLR is used as pretext task so that it can extract features using encoder resnet50 particularly these weights were used in YOLOv5 by changing the architecture of Backbone from Darknet to Resnet 50. By enhancing image quality, our approach significantly improves the performance of object detection in challenging underwater environments. Experimental results on underwater datasets demonstrate that our proposed approach not only enhances image quality but also significantly boosts the accuracy of object detection, addressing crucial challenges in the development of underwater systems.