Research On Object Detection Algorithm Based On Feature Pyramid Network

The task of object detection algorithm is to detect the object category in the scene and locate the object.In recent years,the object detection algorithm based on deep convolution neural network have achieved good detection results and can be applied to a variety of scenes.However,the current algorithm also has many shortcomings,such as in the case of dense and mutual occlusion,the object detection effect is not ideal,and the object has false detection and miss detection.In response to these shortcomings,this paper proposes two improvements on the FPN algorithm.The main work and research word are as follows:(1)Aiming at the problem of dense and occluding objects in complex scenes,using the Inception structure combined with coder-decoder structure in the semantic segmentation model,the paper proposes to use a multi-branch feature network to extract local feature information from the image in the algorithm.Aiming at the impact of the imbalance of the number of positive and negative samples and imbalance of the number of difficult and easy samples on the detection effect,focal loss is used to replace the classification loss function in the FPN algorithm.At the same time,in view of the frequent overlap of objects,occlusion loss is added in the loss function to make the object prediction box and its corresponding real box closely as possible and away from other object prediction boxes.Finally,experiments prove that improved algorithm can improve the accuracy of object detection and the performance of object detection in complex backgrounds.(2)A multi-scale feature fusion object detection network is proposed.First,in view of the limitation of the low resolution of the high-level feature map in residual network,a joint expansion convolution-residual network is proposed to increase the receptive field of highlevel feature map.In order to combine the location information of the shallow network and the semantic information of the deep network,A pyramid pooling network is used to integrate feature information with different depths fully,and a feature pyramid network is used to detect object at multiple scales.Finally,a feature refinement module is proposed to capture the global semantic information to optimize features of each stage in the network.Experimental results show that the network model in the paper effectively improves the detection ability of objects,solves the problem of small object misdetection and missed detection to a certain extent,and enhances the robustness of the algorithm.