Implementation of an Underwater Gesture Recognition and Communication System for Divers Based on Computer Vision

Abstract

Underwater environment factors such as optical attenuation, color distortion, low contrast and dynamic background interference are highly challenging to diver gesture recognition and the current methods have a weakness in their inaccuracy in underwater complex environment. This paper suggests a deep learning-based diver gesture recognition and communication system. Initially, dark channel prior and white balance algorithms are implemented to preprocess underwater images to regain the color of the image and increase contrast. Second, an improved YOLOv5 (You Only Look Once v5) network is built that detects gestures, and a Convolutional Block Attention Module (CBAM) is implemented to increase the ability to extract features. Next, the features of gestures are extracted with ResNet50 (Residual Network 50) and a temporal model is trained with the help of LSTM (Long Short-Term Memory) network in order to identify dynamic gesture sequences. Lastly, a communication scheme is developed to translate the recognition outputs into standard diving commands. Experiments are conducted on a self-built dataset containing 8600 images of 12 commonly used diving gestures. The system achieves an average precision of 96.3%, a recall rate of 95.8%, and an F1 score of 96.0% in clean water environments. Its processing speed is stable at 35 frames per second, meeting realtime requirements and providing effective technical support for safe communication in underwater operations. The system links underwater optical recognition with acoustic modem communication, making it relevant to underwater information-transmission engineering.