Comparing YOLOV3, YOLOV5 & YOLOV7 Architectures for Underwater Marine Creatures Detection

Abstract

The improvement of the marine economy depends on the object recognition of marine species. It is crucial for the intelligence of marine equipment to swiftly and reliably identify marine organisms in a challenging maritime environment. This research report suggests an architecture for a YOLO convolutional neural network trained on aquatic marine life. The most crucial step in object recognition is image pre-processing, which comes before YOLO processing. The pictures produced by the underwater system exhibit symptoms of inadequate lighting, low brightness, and the production of sea fret due to light absorption and dispersion under the sea. The image is improved using the Weighted L1-norm-based Contextual Regularization dehazing technique. This study used a variety of YOLO series to identify seven distinct marine animals, including fish, jellyfish, penguins, puffins, sharks, starfish, and stingrays. When the Enhanced YOLO series is contrasted with the Regular YOLO series, it is found that the former outperforms the latter significantly. Our research also reveals that the YOLOv7 framework performs better in terms of detection accuracy than the YOLOv5 and YOLOv3 architecture while retaining a slower inference time. The enhanced YOLOv7 exceeds the other five frameworks when comparing the YOLO series with enhanced YOLO frameworks, with a mAP of 0.82.