Satellite Image Target Detection And Recognition Based On Deep Learning

Remote sensing is a technique that does not directly contact the target,and observes and acquires target related state information through a detecting device located at a distant distance,and processes the acquired information.Remote sensing technology is widely used in civil and military fields such as ship monitoring,fishery management,environmental monitoring,and enemy military dynamic monitoring.Different types of remote sensing images have different characteristics.For example,SAR images have high resolution and can work all-weather and all-time,but they will contain a lot of speckle noise.Traditional remote sensing detection and recognition algorithms rely on artificially designed image features,which require artificial selection of optimal feature subsets and adjustment of classifier parameters.According to the advantages of end-to-end learning of deep learning,combined with different remote sensing image characteristics,a set of end-to-end SAR image target recognition model and a small target ship detection model combined with multi-scale feature fusion and context information are proposed.The main work and contribution of the dissertation are as follows:(1)A new end-to-end SAR image target recognition model,Flexible Depthwise Separable DenseNets(FDSDenseNet)is proposed.By introducing parameterized activation function and dense convolutional neural network,the recognition accuracy of classification model is improved.A depth separable convolution is proposed to control the model complexity;a variable classification module combining a 1×1 convolution kernel and a classification layer is designed to achieve variable parameter adjustment.(2)Based on the different expressions of different feature channel dimension information streams,sparse connections and group squeeze excitation are used to further reduce the number of parameters.A novel CNN network structure,called group squeeze excitation sparsely connected convolutional networks(GSESCNN)is proposed.The proposed group squeeze excitation module(GSE)is combined with a sparse connected convolutional neural network(SCNN).SCNN uses a unique sparse path connection strategy to reduce the redundancy parameters based on the feature flow.The GSE module adaptively learns the information of different feature channel dimensions by adding additional branches,which improves the model's representation ability.(3)Based on the analysis of existing research work on satellite image target detection combined with deep learning,a new neural network structure,called squeeze excitation skip-connection path networks(SESPNets),is proposed.Based on the framework of two-stage detection,a skip-connection path structure is proposed to enhance the ability of the network to characterize multi-scale features in the target region extraction phase.At the same time,the squeeze excitation module is incorporated into the network to enhance learning of different feature channel information.The ROI align operation and Soft-NMS post-processing are used to further improve the detection accuracy.