Application Of YOLOv3 In Ship Image Detection

In the process of maritime target detection,AI-based target detection has become an essential and important realization tool.In broad waters or port docks,whether of dense aggregation or loose distribution,we need rapid detection and positioning of ships,further classification and segmentation,and the stability and efficiency of target detection is becoming more and higher.The main research content of this paper is to improve the network structure of the YOLOv3 target detection algorithm to improve the network accuracy and recall.This paper takes the water ship image as the research object,conducts deep learning neural network model research,ship image enhancement research,ship image detection training and analysis,etc.The main contents are as follows :1.For the production of the water ship target dat as et,it firstly classifies the DOTA dat as et obtained on the network and selects the ship and other images;studies the ship image blur,dim color,the image enhancement method of the background map,and finally increases the data diversity and the number of targets on the single image to improve the detection effect.2.analyzes the advantages and disadvantages of SSD、Faster-RCNN and learning in YOLO series from the theory and objective of network model.As the key link of training in neural network,loss calculation introduces GIOU loss based on YOLOv3 model and SPP structure,YOLOv3 SPP con vol u tional neural network model is proposed.The YOLOv3 SPP network structure is divided into three parts: Darknet-53、SPP module and prediction feature layer.Through the training and detection of the above ship image data set by three network models,a series of results obtained from the experiment are used to analyze the YOLOv3 SPP network characteristics.The model can complete the task at a time,fast and efficient.