Improved real-time target detection algorithm for similar multiple targets in complex underwater environment based on YOLOv3

Abstract

In this paper, we propose an improved real-time target detection algorithm for similar multiple targets in complex underwater environment based on YOLOv3, which is used for real-time detection of marine organisms (such as holothurians, echinus, scallop and starfish). In traditional methods, the average precision of the classifiers which need to be designed manually is low. In contrast, the target recognition system based on deep learning has great advantages in accuracy and generalization. However, due to their special shapes and appearances, marine organisms are not only difficult to be distinguished from each other, but also easy to be confused with the seabed background. Therefore, the existing target detection algorithms based on deep learning has not achieved good results in the recognition of underwater complex environment. In this regard, we processed the dataset - by data augmentation, image enhancement, sample equalization, which reduces the impact of the dataset on the training results and ensures the accuracy of recognition to some extent. We also made some improvements on yolov3, including adding the feature clear layers and pooling layers, so as to extract more effective features, and also avoid the over-fitting problem in the training process. The experimental results show that our method effectively improves the accuracy of target detection, and has almost no impact on the real-time performance of the system.