Research On Object Detection Technology Of Remote Sensing Ship Image Based On Key Points

In recent years,with the rapid development of satellite remote sensing technology and deep learning technology,detecting specific targets in remote sensing images based on deep learning methods has gradually become a current research hotspot.China has vast sea areas and many important ports,and conducting sea surface target detection has significant value and significance in both military and civilian applications,such as fisheries management,port traffic services,and maritime patrols.The main objective of this article is to study the inaccurate classification and positioning of remote sensing ship image processing on the sea surface under the interference of complex marine environments and various cloud and fog weather conditions,relying on existing deep neural network methods.The main tasks are as follows:(1)Research and analyze the current mainstream object detection algorithms and classify and summarize the main algorithms of object detection.Due to the limitations of anchor-based object detection algorithms,which require pre-set anchors and can result in slow detection speeds,this thesis selects keypoint algorithms that are more real-time to detect various ships under complex sea conditions.(2)In order to solve the problem that too many parameters of CenterNet algorithm and too deep network lead to too slow reasoning speed,we improved its original backbone network by changing the order of increasing and reducing dimensions of residual blocks in Res Net-50,and then introduced depth separable convolution to replace ordinary convolution,so as to reduce the number of parameters and speed up detection.In addition,we have replaced the deconvolution originally used for upsampling with deformable convolution to further improve the fitting ability of the network.By improving the original CenterNet algorithm,we greatly improved the fitting ability and detection accuracy of the model.The experimental results showed that compared with the original method,the average accuracy of the improved method increased by 4.36%,and the Fps value of image processing speed also increased by 14.(3)In response to situations where ships on the sea may encounter severe cloud and fog obstruction,large waves,and dense ship arrangements,as well as the insufficient extraction of remote sensing ship image information by the original feature extraction network,which ultimately leads to low detector recognition rate,this article conducts a parallel operation on the original Res Net-50 network to extract more useful feature information and obtain a more excellent backbone network.In addition,In order to enhance the correlation of remote sensing target features in channel and spatial dimensions,the CBAM attention mechanism was introduced,resulting in an improved D-CenterNet algorithm.Experimental results have shown that the improved D-CenterNet detection algorithm can detect ships annotated in the dataset under complex sea conditions,with an average accuracy of 92.22%,greatly improving the accuracy of the detector.