Construction of Marine Target Detection Dataset And Target Detection Methods

This paper aims to investigate the construction of a marine target detection dataset and target detection methods, focusing specifically on marine ships. To begin with, a collection of marine ship images, 41,162, is obtained through difference resources The LabelImg software is then utilized to visually select the dataset, generating the target object calibration information in the image files. This approach enhances the efficiency of dataset file creation. The dataset is converted from Pascal VOC to YOLO and COCO format. Furthermore, existing target detection algorithms are thoroughly examined, analyzing their structural characteristics as well as the strengths and weaknesses observed in practical applications. Following the preparation of the environment and dataset processing, six detection methods, YOLOv7, Faster R-CNN, RetinaNet, FCOS, CenterNet, and ATSS are selected for training and testing using the constructed marine ship dataset. The evaluation metrics of accuracy, recall rate, and average accuracy are employed to analyze and summarize the experimental results.