Detection And Fine-grained Classification Of Optical Remote Sensing Ship Images

The research on ship detection and classification task in the field of optical remote sensing is a hot topic at present.It plays a vital role in anti smuggling,maritime traffic control,maritime rescue and other activities.With the development of high-resolution remote sensing technology,a large number of data have been produced.A large number of scholars have conducted in-depth research on this problem with the help of deep learning.However,due to the complex and diverse background,intra class similarity and inter class difference,data imbalance and other problems,ship detection and finegrained classification are relatively difficult.In order to achieve fine classification while accurately locating ships,this paper focuses on ship detection and fine-grained classification of optical remote sensing images,as follows:Firstly,aiming at the problem of ship fine-grained classification,which is less studied at present,an algorithm based on attention cutting classification is proposed.This method optimizes the input from the original image into three parts: random cutting image,attention cutting image and original image.Different parts pay attention to different features.At the same time,an attention classification module is designed,which is similar to human vision.The weight of different information in the feature map is allocated,and the high-order features obtained by bilinear pooling improve the discrimination of the model to the target category.The experimental results show that the algorithm can get the best effect on the existing open ship fine-grained classification data set,and verify the robustness on the ship images of other remote sensing data sets.Secondly,due to the lack of ship detection and fine-grained classification data,this paper establishes the relevant data set fgsr-20.There are 14 types of data sets related to military ships and ships,and they are rich in data sets related to commercial ships.The five parameter method is used to mark the ship manually with rotating frame,accurately locate the position and contour of the ship,and clarify its fine-grained label.Thirdly,for the fine-grained detection problem based on rotating frame,the related algorithms are designed.Analyze and solve the difficult problems in ship rotation detection: The problem of data imbalance is solved by using virtual generated data enhancement,balanced feature pyramid and Io U-balanced sampling;Cascade prediction module and rotation alignment convolution are used to solve the problem of feature misalignment.In addition,the fine-grained detection problem is optimized combined with the attention classification module.The experimental results show that the above methods can get good results on the self built data set,and realize the accurate positioning and reasonable fine-grained classification of ships.