A deep learning approach for underwater fish detection

Abstract

Identifying moving objects in video sequences is crucial for various applications, including underwater surveillance, biomedical detection, threat identification, defence, and navy. When comparing images and videos captured in an oceanic environment to those captured in an air, the physical properties of the water medium usually cause degradation, leading to unstable or lost features. Although there are many applications of underwater object detection, researchers find it difficult to extract objects from underwater images due to problems like water body turbidity, blurring, and low image quality. In static background conditions, the background remains stationary, while in dynamic background conditions, both the background and foreground exhibit motion, making it difficult to differentiate between them. The differentiation of both the background and the foreground object is very difficult in dynamic as compared to static. Therefore, the suggested system offers a deep learning-based solution for underwater fish detection that employs three models: YOLOv3, SSD Mobile net v2, and Faster R-CNN ResNet50, all of which were trained on a bespoke dataset called Fish4knowledge. The algorithms have been taught to recognize and reliably pinpoint fish species in underwater photos and videos. To improve performance, data pre-treatment, model selection, and hyperparameter adjustment are carried out. The best-performing model is chosen after evaluation on a different validation dataset. Model updates and adaption to changing undersea conditions are required for long-term accuracy and performance enhancement. The precision value for the Faster R-CNN ResNet50, YOLOv3 and SSD MobileNetV2 is 45.06%, 79% and 98.21% respectively. The research results demonstrate that the SSD MobileNetV2 model gives the highest precision value as compared to the YOLOv3 and Faster R-CNN ResNet50 models.