Research On Densely Packed Object Detection Algorithm Based On Label Assignment And Attention Mechanism

Object detection is an important research area in computer vision,which plays an important role in areas such as face recognition and autonomous driving.With the continuous improvement of the performance of object detectors,applications of object detection in real life are becoming more and more extensive,and the application scenarios will become more and more complex.In dense scenarios will also become,most of the anchor boxes are negative samples,this phenomenon will lead to an imbalance between positive and negative samples,resulting in the problem of unstable network training.If objects are occluded from each other,the features of the occluded objects are incomplete,making it difficult for the model to distinguish objects,this problems increases the difficulty of detection.This paper analyzes these challenging problems in dese object detection in depth.This paper uses label assignment,attention mechanism,non-maximum suppression and other methods to improve the performance of the object detection network and weaken the problems caused by the excessive density of objects to be detected.This main contribution of this paper is summarized as follows:(1)For the problem of positive and negative samples caused by a small-scale state under dense conditions,a adaptive label allocation strategy is designed.During the training process,the threshold is dynamically adjusted according to the localization and classification ability of samples,so that the network can reasonably divide positive and negative samples according to sample density.(2)Aiming at the missed detection caused by occlusion state under dense conditions,the attention mechanism is firstly introduced to solve the problem of lack of characteristic information of the occluded objects by adapting the weight of each element in the feature map through network learning,so as to improve the robustness of the network.(3)a similarity nonmaximum suppression algorithm is proposed,the new method aims to solve the problem that traditional non-maximum suppression with the idea of weight decay,so as to improve the recall rate of the network.Finally,a dense object detection framework(La Net)based on label assignment and attention mechanism is proposed,which connects the three optimization ideas above and combines them onto a object detection network.Through experiments on two general datasets,PASCAL VOC and MS COCO,the detection performance of La Net in dense objects is proved by analysis,and the effectiveness of each module is proved by ablation experiments.