Object Detection Algorithm Based On Sample Selection And Attention Mechanism

With the improvement of computer hardware performance and the development of artificial intelligence,the research of object detection has made great progress and has been gradually applied to various fields.Now the shadow of object detection technology can be seen in transportation,military and agriculture.However,the field of object detection still faces some problems,such as poor detection effect of small targets and easy interference by noise in the detection process.This thesis makes an in-depth study on these two problems and puts forward relevant solutions.The main work of this thesis is as follows:(1)Small object detection algorithm based on sample selection.In view of the poor effect of object detection algorithm on small object detection,this thesis proposes an fe-ssd model based on SSD model.The feature enhancement and feature fusion module is added to the FE-SSD.Through the feature enhancement module,the receptive field size of the shallow feature layer is expanded and the feature context information of small target is effectively extracted,while the feature fusion module can integrate the semantic information of the deep feature layer into the shallow feature layer.At the same time,At the same time,this thesis proposes a dynamic training sample selection algorithm.The traditional way of dividing positive and negative samples by fixed threshold leads to less positive samples of small target matching,so the ability of small target detection is poor.The dynamic training sample selection algorithm proposed in this thesis improves the number of positive samples of small objects.Experiments show that the proposed algorithm can effectively improve the detection accuracy of small objects.(2)Object detection algorithm based on attention mechanism.Aiming at the problem that the object detection algorithm is easy to be disturbed by noise in the detection process,this thesis designs a hybrid domain attention model.By dividing the feature map into multiple subspaces and learning the individual attention weight of each subspace separately,it can not only learn cross-channel information,but also learn multi-size and multi-frequency features.Adding the attention module to the object detection model can effectively reduce the interference of noise information,make the model pay more attention to the key feature information of the detected object,and improve the detection performance of the object detection model.Experiments on FE-SSD model show that the hybrid domain attention model proposed in this thesis can effectively improve the detection accuracy.