Research On Detection Of Floating Objects On Water Surface Based On YOLOv3

Floating objects such as water surface garbage,flooded aquatic plants and water blooms pose a serious threat to the safety of the water environment,and timely detection of surface floating objects so as to salvage and clean them is an important measure to build a clean and safe water environment.Surface float detection is a deep integration of intelligent monitoring technology and water surface environmental supervision business,but the current research in this direction mainly focuses on the monitoring of scattered individual floating objects.In addition,a learning rate decay scheme design algorithm based on exponential function is proposed,which aims to quickly find a more suitable learning rate decay value to make the model converge,and determine the learning rate decay node and training period..The process of designing the learning rate decay scheme for the YOLOv3 model using this algorithm is demonstrated in detail through several sets of comparison experiments,and the obtained learning rate decay scheme is compared with that using the cosine annealing decay strategy.The experimental results show that the loss functions of the YOLOv3 model using the former can all converge,which verifies the effectiveness of the proposed algorithm,and the mean average precision（m AP）obtained from the former experiment is higher than that of the latter with the same other parameters.In order to solve the problem that the YOLOv3 model is fast but slightly low precision,we propose a detection model S-YOLOv3 that incorporates the SE attention module in the channel dimension of the YOLOv3 model,which uses the K-means++ algorithm instead of the K-means algorithm to cluster the labeled bounding box size of the dataset and reduce the negative impact caused by the random selection of the initial clustering center.In addition,the localization loss in the loss function of the YOLOv3 model is improved,and GIo U Loss is introduced to improve the localization accuracy.The experimental results show that S-YOLOv3 outperforms other models in MW-Float dataset compared to YOLOv3 and other commonly used models in the field of water surface target detection,and the m AP reaches83.1%,which can basically meet the demand of water surface floating object detection.