Research On Underwater Target Detection Method Based On Deep Learning

Underwater vision is an important research direction in marine studies.Due to the uncertainties in the underwater environment,there is a problem of class imbalance in the obtained underwater images.Moreover,the water contains a large amount of solvents,particles,and other heterogeneous media,resulting in less light entering the camera compared to the natural environment.This leads to blurring and distortion in underwater imaging,resulting in low accuracy and false detections in underwater detection.To address these issues,this research focuses on three main aspects:To tackle the problem of class imbalance in underwater images,a method based on generative adversarial networks(GANs)is proposed for underwater small-sample augmentation.By collecting a small number of underwater target images,a underwater dataset is built.An unsupervised image-to-image transformation method is introduced,which includes a GAN architecture and incorporates attention modules and a new normalization function called A-L.Multiple styles of underwater images are synthesized using unpaired input and real targets.To improve the low accuracy issue in underwater detection,a YOLOv7＿UW underwater object recognition model is proposed.L1 regularization is applied before the batch normalization(BN)layer to remove layers with smaller impact factors through sparse training.Coordinate attention mechanisms are introduced between the neck network and prediction network to enhance detection accuracy.The model achieves both high speed and high accuracy in underwater target detection while being lightweight.A series of experiments were conducted,and the results showed that the proposed algorithm can accurately detect various marine organisms with higher accuracy than YOLOv7 and YOLOv7 x models.By comparing the performance of student behavior detection in different scenarios,the improved algorithm achieved an average accuracy of 90.6%,a recall rate of 79.3%,mAP@0.5 of 86.9%,and mAP@0.5:0.95 of 54.1%,all superior to the compared algorithms.To meet practical application and operational requirements,a complete underwater object detection system is designed using the YOLOv7＿UW underwater detection model.This system is based on the PyQt5 application framework and integrates data import and result display into the user interface.It enables real-time detection of underwater images or videos,meeting the needs of practical engineering.