Object Detection Method Based on PVTv2
Yuzhe Chen, Zhe Min Zhuang, Cheng Chen · 2023
Ship detection is the first barrier to ensure the safe operation of ships, to ensure that there are no serious technical loopholes or safety problems in the process of navigation. With the progress of remote sensing technology, the use of deep learning as a method for target detection has become a hot topic. Compared with the traditional manual detection method, the target detection algorithm based on deep learning has the following advantages of no need for manual feature design, good feature expression ability and excellent detection accuracy. At present, most of the methods used for ship detection are traditional methods, which are not robust to illumination, morphological direction changes, occlusion and so on. In this paper, three improvements are made to solve the above problems. Firstly, Transformer backbone network — PVTv2 is adopted. PVTv2 uses convolution to extract local continuous features, position coding with zero-padding, and attention layer with linear complexity of mean pooling, which can extract features more fully. Then, the width and depth of the pvt are increased to further extract features; Finally, the Cutout technology is used, which can improve the robustness and overall performance of the convolutional neural network, and improve the generalization ability of the model. Experimental results show that, compared with the original model, the average accuracy mean (mAP) is improved by 4. 39% on the HRSC2016 dataset, the validity of the model is proved.