Deep Learning-Based Fish Detection in Turbid Underwater Images

Abstract

The successful results of the deep learning models in many areas have been the exit gateway to the problems faced in the challenging conditions of underwater studies. One of these problems is the detection of fish in images with a high turbid and background noise. Therefore, the detection of fish in turbid and background noisy water is an important threshold to be overcome to classify them and track their paths. In this study, videos were taken from the reservoir basin in Kahramanmaraş Ceyhan region with two different cameras. Then, a novel data set is presented which contains 400 images for the detection of fish in the wild. By using these data set, the state-of-the-art detection models, YOLO-V2, YOLO-V3, YOLO-V3 Tiny and MobileNet-SSD networks are trained with fine-tuning strategy, and then they are compared over the precision, recall and mean Average Precision (mAP) performances.