The Research Of Ground Object Recognition Algorithm Based On Lightweight Convolution Network

Ground object recognition and detection is of great significance for public security,ground reconnaissance and traffic dispersion.However,few features can be extracted from these small objects.Therefore,how to improve the recognition ability of small objects and reduce the calculation amount has become a research focus.In view of this issue,following aspects of work is mainly completed:First,commonly used object detection frameworks are analyzed and Lighthead-RCNN is selected and improved.Better prior anchors are generated from the datasets,which reduces the difficulty of training optimization and improves the detection accuracy;a spatial attention module is used to optimize the distribution of foreground and background features;PSRoiAlign is used during ROI pooling to improve the distortion and misalignment caused by rounding during candidate box pooling.The improvement of the algorithm accuracy is verified through the comparison experiment on the open dataset.Then,a lightweight backbone network Dense Residual Shuffle Net(DRSNet)is proposed.This backbone not only has the residual structure from Res Net and dense connection from Dense Net,but also improves the structure of the feature fusion on the basis of FPN(Feature Pyramid Networks)to fuse different feature maps more adequately.Then the basic module in Shuffle Net V2 is selected as the lightweight convolution module of the network,which reduces the amount of computation and parameters of the network.The advantages of this backbone network over other networks are verified by experiments and quantitative analysis.Experiments on different datasets prove the robustness of the proposed algorithm and its superiority over other algorithms.Finally,the accuracy of the algorithm is further improved for the problems of unbalance and occlusion.After using hard negative example mining to relieve the unbalance between background and foreground examples,the balanced sampling is used when sampling the candidate boxes generated by RPN to relieve the unbalance between different foreground categories,which increase the sampling probability of small quantity categories.Secondly,to improve the low recognition precision of obscured objects,Soft-NMS is used,which decreases the scores according to the Io U with the highest score box,so as to help distinguish between false positive and ground truth.Through the comparison experiment between algorithms before and after the improvements,the effectiveness of the improvements under the unbalanced problem and the occlusion problem is verified.