Small Object Detection Method Based On Deep Convolutional Neural Network

In recent years,the field of object detection has developed rapidly.As the focus and difficulty of object detection,small object detection has gradually attracted many researchers.In real life,there are many scenes with a large number of small objects,such as pedestrians photographed by drones,vehicles and traffic signs in the distance during automatic driving.Therefore,small object detection requires both accuracy improvement and real-time performance of the algorithm.Aiming at the problems of small object with small pixels and insufficient feature information,on the basis of the single-stage object detection method YOLOv5,this thesis proposes a method that is more conducive to small object detection.While improving the detection accuracy of small objects,the proposed method has a certain real-time performance.The main work of this thesis includes the following points:1.Starting from the single-stage object detection method YOLOv5,this thesis introduces its main components.In addition,this thesis analyzes the YOLOv5 network structure in combination with common algorithms for small object detection,and improves and experiments some of its components.Finally,through experimental analysis,it is concluded that some components of YOLOv5 have deficiencies in small object detection,which provides a theoretical basis for subsequent improvements.2.In order to solve the problem that the existing object detection methods are not effective in small object detection and most of them cannot achieve real-time processing,this thesis proposes a small object detection method based on parallel convolution and fast multi-scale fusion.In terms of feature extraction,an adaptively spatial parallel convolution module(ASPConv)is proposed to alleviate the lack of spatial information of objects and adaptively obtain the corresponding spatial information through multiscale receptive fields,thereby improving feature extraction capability.Additionally,Split-Fusion submodule is proposed to effectively reduce the time complexity of ASPConv.In terms of feature fusion,a fast multi-scale fusion module(FMF)is proposed to alleviate the insufficient fusion of both semantic and spatial information.FMF module uses two fast upsampling operators to first unify the resolution of the multi-scale feature maps extracted by the network and then fuse them,thereby effectively improving small object detection capability.Comparative experimental results prove that the proposed method considerably improves the accuracy of small object detection on multiple benchmark datasets such as Tiny Person,Tsinghua-Tencent 100 K,UAVDT and MS COCO,and achieves high realtime performance.