Research On Object Detection Based On Atrous Convolution And Edge Guidance

As an important part of computer vision,object detection is widely used in daily life.The current mainstream target detection algorithms have various detection forms,among which RPDet(Rep Points Detector),as a single-stage detection model,has the characteristics of fast detection speed and good detection accuracy.However,with the increasing complexity of target detection tasks,the detection accuracy of RPDet model can no longer meet the application requirements.In this paper,the RPDet model is used as the benchmark model to improve its feature extraction network,feature fusion network,detection header network and post-processing stage,so as to achieve the purpose of improving detection performance.Improvements to this article are as follows:(1)In order to solve the problem of missing detection caused by single field of feature extraction and insufficient feature information fusion for targets of different scales in the process of target detection,this paper proposes a feature extraction and fusion method based on atrous convolution,which improves the feature extraction network and feature fusion network on the basis of the reference model.At the stage of feature extraction,Deformable Atrous Convolution(DAC)module is proposed and introduced into the feature extraction network to improve the feature extraction ability of the model.At the Feature Fusion stage,Multi-Scale Atrous Feature Aggregation(MSAFA)and Layer Fusion(LF)modules were proposed.In Feature Pyramid Networks(FPN),MSAFA module is introduced to synthesize multi-scale output features of branches at different levels,and LF module is introduced to realize the fusion between high and low layer features of FPN networks.This method solves the problem of feature extraction and fusion of different scale targets in the process of multi-scale target detection and can effectively improve the performance of detection model.(2)In order to solve the problem of incomplete target recognition in the case of complex background or overlapping objects,this paper proposes an optimization method of target detection based on edge guidance.By adding edge auxiliary branches,edge verification and regression branches of the benchmark model are fused together,and the prompt information containing edge verification is used to guide target detection tasks.The problem of incomplete object recognition in the case of complex background or object overlap is solved.In the network post-processing stage,the Confidence Propagation Cluster(CP-Cluster)algorithm is introduced to eliminate the low quality prediction frame and enhance the true positive propagation frame.This method solves the problem of poor detection model performance caused by the failure of the RPDet model to utilize the edge information of objects and the relationship between prediction boxes,and improves the detection performance.Finally,the ablation experiment was carried out on the data set with the proposed method,and the comparison experiment was conducted with the current advanced detection model.The experimental results proved the effectiveness of the improved model,and the detection accuracy was significantly improved.