Convolution Neural Network for Identification of Underwater Objects

In this paper, we propose an approach to creating an automatic monitoring system, based on convolution neural networks to perform real-time underwater object identification and classification. In this case, we used self-collected dataset of underwater objects in several areas of the Black Sea and the Sea of Azov, which was preprocessed and augmented using Keras. For object recognition and identification, we used Darknet YOLOv4. Despite all environmental difficulties, we can reach mAP@0.5 = 72,22%, with range of accuracy per class from 85.78% to 100 %. Thus, the proposed method can be useful in fisheries and aquatic farming, underwater engineering activities, etc.