Research On Target Detection Algorithm Based On Long Tail Distribution Data Set

With the rapid development of deep learning technology,more and more applications of artificial intelligence appear in daily life.The representative technologies,such as face recognition,pedestrian recognition and license plate recognition,have created great convenience for human life.In these applications,the object detection algorithm plays an important role.Although the existing algorithms achieve high accuracy in tasks with fewer categories,with the increase of the number of categories,their performance will deteriorate dramatically,which makes it unable to meet the algorithm requirements in complex natural scenes such as automatic driving,robot inspection and so on.Based on the analysis of large-scale multi classification dataset LVIS,the longtail distribution of dataset is the main bottleneck of multi classification target detection.Longtail distribution refers to the phenomenon that the number of each category in the training set is extremely unbalanced and the rare category is not fit.In recognition field,"reweighting" and "resampling" are usually used to deal with such problems.However in the detection task,the instance and the picture are coupled with each other,and it is difficult to distinguish the rare class from the background,which brings greater challenges to the detection of the rare class.Based on Faster RCNN,a new loss function and sampling strategy are proposed.In the aspect of loss function,the competition among classes is restrained by splitting foreground background loss,and the gradient proportion of training is increased by gradient protection and reweighting of rare classes.In the aspect of sampling strategy,first,Top K anchor increases the number of samples of rare anchors in the RPN stage,thereby increasing the scores of rare proposals,and then improves the number of rare class samples through RCNN sampling of adaptive Io U threshold,which effectively improves the precision of rare class with image level sampling of instance priority.The experiment of this paper is based on LVIS dataset.Compared with Faster RCNN,the proposed method has an overall improvement of 4.9% in m AP and an improvement of 13.3% in rare class m AP.The proposed method is also better than other methods in the same field.Further experiments prove that this experiment has achieved good results in different model structures and training settings,showing good practicability and generalization,and provides a good example for related technology research and practical application.