Ship detection based on Improved YOLO Algorithm

With the continuous development of marine economy, marine ship detection has a very wide application prospect. To solve the problems of low ship detection accuracy and easy missing detection of small targets in complex marine environment, an improved YOLOv3 target detection algorithm is proposed. Firstly, before data training, the methods of image blending and small target enhancement are adopted to overcome the influence of complex environment background on target detection and the imbalance of small target data, and improve the network's ability to extract image features and detect small targets during the training process. Then, the convolution in the original network is replaced by Octave convolution, which reduces the amount of calculation during network training and improves the detection speed of the network through image frequency division. At the same time, GIOU loss function is introduced to alleviate the problems that the gradient is zero when the detection frames do not overlap, and IOU cannot distinguish the alignment of prediction frames and annotation data in different ways, thus improving the accuracy of detecting the target position. The experimental results show that the mAP of the improved YOLOv3 algorithm on the self-made ship data set can reach 84.07%, and the improved algorithm is improved by 3.52% compared with the original algorithm, and the detection speed can reach 31.8 frames/s, which effectively improves the accuracy and speed of detecting ships at sea.