Abstract: In underwater target recognition,acoustic data is greatly affected by noise and susceptible to interference,while the difficulty of obtaining optical data increases significantly with depth.This paper proposes a method of fusing acoustic and optical images to improve the accuracy of underwater target recognition.To address the scarcity of corresponding acoustic and optical image datasets for underwater targets,a Cycle-GAN network is employed for dataset sample augmentation,followed by image enhancement processing on the generated dataset.In the field of target recognition algorithms,the Transformer cross-modal attention module and the channel prior convolutional attention mechanism are integrated into the YOLOv8 algorithm to improve target recognition accuracy and precision.This study utilizes a sonar-scanned shipwreck target dataset.Experimental results indicate that to the cross-modal acoustic-optical fusion target in target recognition algorithms,the average and average accuracy rates have respectively improved 0.175 and 0.165. Constructed optimized backbone network enables cross-modal integration of acoustic and optical features,enhancing the efficiency of feature extraction and addressing the challenge of distinguishing the edges of submerged shipwrecks from their surrounding environments.

Key words: target recognition, acoustic-optical fusion, Cycle-GAN, cross-modal attention, YOLO algorithm, fusion algorithm, Transformer