Research On Remote Sensing Image Of Sea Surface Object Detection Based On DNN

In recent years,as an important marine vehicle and a valuable military target,ships have important practical significance for the automatic detection of their remote sensing images.However,in reality,remote sensing images are often affected by complex sea conditions,such as cloud cover,shore-based interference,etc.These factors will increase the complexity of ship detection.Based on the study of the characteristics of remote sensing images and relying on the existing deep neural network method,this paper carried out the research of DNN-based remote sensing image sea object detection technology.The main research contents include the following:(1)Researched the application of deep neural networks to the detection of sea surface objects in remote sensing images,and used Faster R-CNN-vgg16,Faster R-CNN-res101,YOLO v3 algorithms to complete the network training.The data set of the sea object is collected and labeled by itself,which ensures the number of training samples,enriches the orientation of the ships in the sample,and makes the learned model extensive.(2)Aiming at the problems of low detection accuracy and long time consumption of Faster R-CNN algorithm,this paper proposes an improved algorithm based on Faster R-CNN.First,use the clustering method to regenerate the size and number of RATIOS.Second,replace NMS with Soft-NMS.Experiments show that,compared with Faster R-CNN-vgg16 and Faster R-CNN-res101,the m AP value of the improved Faster R-CNN algorithm is increased by 2.02%and 0.94%,respectively.(3)Aiming at the problem of low accuracy of the object detection model of YOLO v3 algorithm,this paper proposes a new improved algorithm.First,the Kmeans++ clustering method was used to analyze the data set,and an anchor box value suitable for the data set was generated.Second,the new detection algorithm was obtained by modifying the feature extraction network structure and using GIo U as the loss function of coordinate prediction.Experiments show that the G-YOLO v3 method has significantly improved detection accuracy and speed.Compared with Faster R-CNN-vgg16,Faster R-CNN-res101,improved Faster R-CNN and YOLO v3,their accuracy values ? ? have been improved by 7.86%,6.78%,5.84% and 1.38%,respectively.In terms of detection speed,the G-YOLO v3 method detects each image at a speed of 0.115 s.Compared with the above four methods,the detection speed is increased by an average of 0.138 s.In conclusion,the sea surface object detection method proposed in this paper can effectively improve the object detection effect under complex sea background,and provides a better solution for the detection of sea surface object.