Research On Object Detection Algorithm Based On Switchable Atrous Convolution And Attention Mechanism

Object detection is designed to classify and locate targets in an image.In recent years,the target detection algorithm based on deep learning has performed particularly well and has been applied to human life,such as driverless cars,face recognition in the railway sector and intelligent security in residential areas,which show that target detection algorithms play an increasingly important role in improving human quality of life and maintaining social order.However,due to the insufficient extraction of target information and the interference of the background environment,there are still false detection and missed detection of target detection,and how to detect the target more accurately has become a challenging problem.Therefore,in view of the above difficulties,this article combines Switchable Atrous Convolution(Switchable Atrous Convolution,SAC),Global Context Network(Global Context Network,GCNet),Path Aggregation Feature Pyramid Networks(Path Aggregation Feature Pyramid Networks,PAFPN),and Balanced Feature Pyramid(Balanced Feature Pyramid,BFP)proposes improvements.The main research work is as follows:(1)Aiming at the problem that the detection accuracy of the object detection algorithm is low,we propose a target detection model based on the residual network Res Net50 and switchable atrous convolution SAC.The SAC structure acts the switching function on the deformable convolution with a void rate of 1 and the deformable convolution with a void rate of 3,respectively,and fuses the feature information after the action,expands the feature receptive field and realizes the adaptive selection of the feature receptive field,and enhances the model’s ability to extract image feature information.Experimental results on the Pascal VOC dataset show that the overall detection accuracy of the model and the detection accuracy of various types have been effectively improved.(2)Aiming at the problems of target miss detection and error detection caused by the interference of the image environment,we propose a target detection model that integrates the attention mechanism of GCNet.The GCNet attention mechanism belongs to the global attention,and helps the model focus on the target information from the global perspective by generating the global attention feature map,so as to avoid the situation of target leakage and error detection caused by the interference of the picture environment to a certain extent.Experimental results on the Pascal VOC dataset show that the model has improved detection effect when detecting disturbing targets in the picture environment.(3)Aiming at the problems of low utilization of low-level features and unbalanced utilization of multi-scale features of the object detection model,we propose a target detection model based on PAFPN multi-scale feature fusion and BFP feature enhancement.PAFPN helps the model to enhance the use of low-level localization information,and BFP further enhances multi-scale features and balances multi-scale feature information.The detection accuracy of this model increased to 83.7% in the Pascal VOC dataset,and 7.2%higher than that of the original Adaptive Training Sample Selection(Adaptive Training Sample Selection,ATSS)algorithm in the MS COCO dataset,indicating the effectiveness of the proposed model.