Target Detection Algorithm Based On Feature Pyramid Structure

Object detection,as a research hotspot in the field of computer vision,is also one of the most important and challenging basic tasks in the field of computer vision.It plays an important role in the fields of autonomous driving,security monitoring,defect detection and other fields.In recent years,through the theoretical research of a large number of scholars,there has been considerable progress in the field of target detection.However,there is still room for improvement in the detection of small target objects,the adaptability of the network to scale changes,and the ability of the network to express features.Aiming at the above problems,this paper analyzes the basic principles of the two-stage target detection algorithm based on the feature pyramid structure,improves the feature extraction network structure and the feature fusion structure,and proposes a Deformable Multi-scale Feature Perception Network(DMFPN)with strong robustness to the target scale,scene change and boundary deformation.The main work is as follow:(1)Deformable Multi-scale Feature Perception Network.Aiming at the small target problem and the basic requirements of target detection tasks,it provides a higher resolution feature map for the subsequent detection tasks.Based on Resnet50 network,this paper improves the structure of residual bottleneck block and reduces the pooling times of the whole network,enhances the detail information,and ensures that detection tasks are performed on higher-resolution feature maps.Deformable Multi-scale Feature Fusion Module is designed,which can extract and fuse multi-scale features.It can process the feature map of the pyramid structure,and improve the robustness of the network to scale change.(2)Target detection algorithm based on two-way feature pyramid structure and joint normalization method.Based on the top-down structure of the original feature pyramid,the bottom-up structure is added,the transfer of detail information is added,and the global receptive field information is used as the guidance information to weight the channel of the low-level feature map,and the up sampling operation involved in the feature pyramid is improved.The bottom layer of the network improves the robustness of the network to color,brightness and style by changing the structure of the normalization layer in the backbone network,combining batch normalization and instance normalization,reducing the appearance differences between different individuals of the same kind,filtering the complex appearance changes,improving the generalization ability of the network output model,and speeding up the model convergence.(3)In order to verify the effectiveness of the algorithm in this paper,the algorithm is verified on the Pascal VOC dataset,and the output of the network structure is visualized.The mean average accuracy of this algorithm is 82.30% in Pascal VOC dataset.Under the objective evaluation index,the structure of this paper is compared with different algorithms quantitatively and the subjective visual effect is compared with the original algorithm.