| Synthetic Aperture Radar(SAR)has the advantage of working all day and all weather.It is of great significance to use SAR target detection technology to detect ground and sea targets in real time.With the development of satellite technology,it is more and more easy to obtain high-resolution SAR images.However,most of the existing SAR image target detection algorithms can not be applied on a large scale.The main reasons are as follows:firstly,the robustness of traditional methods are poor,so the detection accuracy is low.Secondly,due to the limitation of objective conditions,it is difficult to apply cluster GPUs and highprecision complex algorithms on a large scale.Then,although the detection speed of the lightweight algorithms are fast,but the detection accuracy is low.Thus balancing the detection speed and accuracy becomes a problem to be solved.Finally,the accuracy and speed of the SAR image ship detection methods based on deep learning is far superior to the traditional algorithms,but it is still in its infancy and has a lot of room for development.Therefore,in view of the shortcomings of the above SAR image ship detection algorithms,this thesis uses deep learning technology to detect ships in SAR images.The main contributions are as follows:Firstly,a target detection method for SAR image that improves channel attention is designed to solve the problem of low detection accuracy of traditional target detection methods.This method designs an improved channel attention module,which uses the channel mapping method to get channel weights.This module not only takes into account the statistics of each input feature layer,but also calculates the channel weights more easily,which can give more attention to the important channels in the feature layer.Using this module in the model can further improve the detection accuracy and speed.Secondly,a lightweight SAR image target detection method to improve channel attention is designed to overcome the slow detection speed of traditional target detection methods.The method designs an improved feature extraction module and an improved feature fusion module,in which the improved feature extraction module only extracts features once for each dimension of the feature layer,and the improved feature fusion module only uses two effective feature layers for fusion.Using these two modules in the model can further improve the detection speed.Then,a lightweight SAR image target detection methods that combines spatial attention with improved channel attention is designed to solve the problem of low detection accuracy of the lightweight target detection algorithm.This method designs a joint spatial attention and improved channel attention module,which enables more attention to important areas in the feature layer and further improves the detection accuracy when used in the model.Finally,the ship target detection experiment is performed on the dataset SSDD and the performance analysis is performed.The results are as follows:the SAR image target detection method designed to improve channel attention has the highest detection accuracy,the lightweight SAR image target detection method designed to improve channel attention has the fastest detection speed,and the lightweight SAR image target detection method that combines spatial attention with improved channel attention has the best balance in accuracy and speed.The above work is helpful to explore the accuracy,speed and their balance of SAR image ship detection,and contributes to the large-scale application of SAR image ship detection algorithm based on deep learning. |
