An Improved Object Detection Model Based On Feature Enhancement And Anchor Optimization

With the development of convolutional neural network,object detection develops rapidly,and the one-stage detection algorithms have become a great concern of people.Analyzing the problems existing in these algorithms and putting forward corresponding solutions can improve the detection accuracy of the algorithm while ensuring the detection rate.An object detection model based on feature enhancement and anchor optimization is proposed to solve the problem of the lack of feature expression ability and the imbalance and mismatch of the hand-designed anchors.As follows:Aiming at the problem of insufficient expression ability of detection features,information fusion between detection feature maps and feature map refinement are designed to enhance the detection features.Firstly,reverse fusion operations are performed between multiple detection feature layers of different scales to improve the semantic information of shallow feature maps through the fusion of feature information at different levels.Secondly,dilated convolution is used to enlarge the receptive field of shallow feature maps and capture more context information.More global information is fused to the shallow feature maps without losing details.In order to further enhance the representation ability of the object detection model,the detection feature maps of different scales are refined to improve the attention of each detection feature map to its own information.Aiming at the problem of imbalance and mismatch in anchors by hand-designed,the shape of anchors is changed from artificial design to dynamic prediction.The detection feature maps are input into the designed adaptive generation network,and the optimal shape of anchor at each location can be learned adaptively according to the context information of different locations at feature map.After the anchors are adaptively generated,the anchor optimization model is introduced on each detection feature layer associated with the anchors to further optimize the generated anchors,and the anchors with a higher probability of containing the objects and more matching with the locations and sizes of the objects are obtained.The anchor optimization model uses semantic features to identify and remove the negative anchors that are easy to classify to reduce the search space of the objects,and at the same time adjusts the locations and sizes of the anchors preliminarily,providing high-quality anchors for the final classification and regression.Finally,by combining the two improved methods,an object detection model based on detection feature enhancement and anchor optimization is proposed.After the enhancement of the detection features,the adaptive generation and optimization of the anchors are carried out by using the enhanced detection feature.The experimental results show that the combination of the two improved methods is effective and can be applied to the multi-scale detection framework to greatly improve the detection effect.