ABSTRACT Detection of submerged objects in underwater environments is uniquely challenging due to low visibility, variable light conditions, and high noise and distortion in Underwater Aquatic System (UAS). The traditional object detection models, including the most recent versions of YOLO, drop too often in accuracy and robustness under such hostile conditions. To handle these, YOLO-Transformer Hybrid model is proposed which combines YOLO's real-time detection capabilities with enhanced contextual understanding accorded by transformer-based attention mechanisms. The YOLO backbone extracts the essential features from input image and processes the image through a number of convolutional layers for real-time object detection. It adds attention layers to the transformer module, focusing on relevant regions of the image to extract long-range dependencies and contextual relationships possibly missed by convolutional layers in isolation. Then augment this hybrid approach with IoT sensor data to add valued-added environmental context that may improve detection accuracy. IoT sensors can facilitate dynamic adaptation to the variations in underwater conditions by continuously providing real-time data on temperature, turbidity, etc. The proposed model ensures 97.5% detection accuracy, outperforming traditional YOLO models under challenging underwater scenes.
Neelamegam et al. (Fri,) studied this question.