Research On Object Detection Based On Feature Enhancement And Multi-Level Fusion

As a mainstream research direction in the field of artificial intelligence,object detection algorithm is also one of the four basic tasks in computer vision tasks,and plays an important role in mobile payment,traffic safety,medical and health care and other fields.In recent years,with the update of computer hardware and the in-depth theoretical research of deep learning methods,the field of object detection has been rapidly developed,but the robustness of the algorithm and the ability of image positioning still have further room for improvement.Aiming at the above problems,this paper analyzes the basic principle of YOLOv3 algorithm,and improves it in backbone network and feature fusion network.The main work contents are as follows:Firstly,the E-YOLOv3 algorithm based on feature enhancement is proposed to solve the problem of low accuracy and poor robustness caused by insufficient feature fusion of YOLOv3 algorithm in backbone network.The structure uses the trunk and branch network parallel computing approach to feature extraction,and adopt different depth in the branch network structure of convolution module to extract different characteristics,make the network feature fusion more fully,and then,aiming at the problem of branch network to extract the characteristics of coarse to join in the branch network structure refinement of image feature spatial separable convolution module,The detection accuracy of the network is further improved,the image features extracted by the double-branch structure are more comprehensive,and the robustness of the whole network is enhanced.Secondly,the EM-YOLOv3 algorithm based on feature pyramid network structure based on multi-level feature fusion is proposed to solve the problem that some shallow feature information in feature fusion of E-YOLOv3 network is not fully utilized,and the network loses the original image target location information by screening the shallow feature information prematurely.The structure using features integration with superposition approach to enhance the flow of information between the shallow and deep information,and then,according to the characteristics of the multiple level feature fusion deep semantic information pyramid structure with shallow position information imbalances by extended convolution module,make the network layer size under the condition of without sacrificing features increase the receptive field,The detection accuracy of the network is further improved and the ability of the network to locate the image is enhanced.Finally,in order to verify the effectiveness of the proposed network structure,the proposed double branch backbone network structure based on feature enhancement and the proposed feature pyramid structure based on multi-level feature fusion were performed on Pascal VOC data sets,respectively,ablation experiments,comparison experiments and network visualization experiments.The results show that the accuracy of the proposed algorithm reaches 84.44% on Pascal VOC,and the performance of the proposed algorithm is better than the original algorithm in single target detection,multi-target detection,small target detection and occlusion target detection tasks in network visualization comparison.