Reasearch On Small Object Detection Algorithm Based On One-stage Model

Object detection is an essential field of computer vision.As an important branch of object detection,small object detection has excellent application value,such as object detection from the perspective of UAV,flaw detection in industry,vehicle and ship detection in satellite remote sensing images,etc.However,the precision of small object detection algorithms is currently poor.There are still many problems to be solved,such as the lack of information on the small object itself,leading to fewer features available for the model,and the lack of model learning of small object features in multi-scale scenes.Moreover,the robustness of small object detection to object location error is poor,and it is difficult to locate small objects accurately.To improve the performance of small object detection,this thesis focuses on the above issues.At the same time,considering the speed requirements of small object scenes,this thesis selects the one-stage object detection algorithm YOLOv4 as the baseline.Based on the Vis Drone 2019 visible light image dataset,the small object detection algorithm based on the one-stage network is studied.The research contents are as follows:1.Aiming at the problems of small object detection,this thesis studies the small object detection method based on information enhancement.The method includes an improved Mosaic data augmentation method and an improved loss function.The improved data augmentation method increases the size of the small objects in the training process and strengthens the information carried by the small objects.However,it also reduces the proportion of small objects.Therefore,this thesis also introduces a nonlinear penalty based on the size of small objects in the loss function,which can enhance the penalty for small object positioning errors and suppress the negative impact of the improved Mosaic data augmentation method.The improvement based on information enhancement improves the positioning accuracy of small objects and enhances the self-information of small objects.The experiment also proves the effectiveness of this method.2.In view of the difficulty of small object detection,this thesis introduces an attention mechanism to the model.First,this thesis introduces an attention module into the neck network,which can effectively make the network focus on small object regions,therefore making small objects easier to be detected.Experiments show that placing the attention module in the neck network can improve the detection performance of small objects.Secondly,this thesis studies the weighted feature fusion module and the feature layer attention module to introduce an attention mechanism at the feature fusion to improve the model’s attention to small objects.Through the feature fusion scheme,to a certain extent,this thesis solves the problem that small objects have few learnable features and are difficult to locate.Experiments show that the performance of small object detection is improved by introducing the attention mechanism.The experimental results of this thesis fully prove the effectiveness of the above improvements.The improvement research done in this thesis has achieved a 2.8% AP@.5:.95 improvement and a 2.9% AP@.5 improvement on the Vis Drone2019 dataset.The mean average precision of small objects has improved by 1.5 %,and the average recall has improved by 1.9 %.