Application of Neural Network Technologies for Underwater Munitions Detection

Abstract

In the article, the substantiated proposals for the use of YOLO family neural networks to detect the underwater undetonated munitions are proposed. At the same time, the YOLO3, YOLO4 and YOLO5 neural networks previously trained on the MS COCO dataset are used. The retraining of YOLO3 and YOLO4 neural networks is carried out on the modified Trash-ICRA19 underwater trash dataset, with the number of object classes equal to 13 and 2 of them are fictitious. The average class detection accuracy of 13 object classes using YOLO4 in the mAP50 metric is equal to 75.2 or 88.9% taking into account fictitious classes. The images obtained from video recordings of the demining reservoirs process with the help of remotely operated underwater vehicles (ROV) are used to test neural networks. The improved neural network as a cascade of several serially connected YOLO-segments with multi-pass image processing and tensor-matrix description of the attention mechanism are proposed. The recommendations for further increasing the efficiency of the neural network method of underwater munition selection are developed.