Research And Application Of Small Target Foreign Object Detection Based On YOLOv4

This paper will design and propose a fast localization and detection for small target foreign objects based on deep learning target detection algorithm,and at the same time make a self-made data set about small target foreign objects on airport runways,based on which target localization and detection will be performed based on Open CV platform in order to be better ported to hardware devices such as video surveillance in the future.The main work includes.First,the backbone feature network CSPDarknet53 is replaced with a lightweight Mobile Netv2 network,which is a lightweight network proposed by Google.The internal structure of the network uses a deep separable convolution and an inverted residual structure to make the algorithm scale lightweight while maintaining a high detection accuracy.Second,based on the improved YOLOv4 algorithm,an effective feature layer with a detection head size of 104×104 is added to extract the stronger position information from the shallow features,and then the 104×104 feature layer is fused with the other three feature layers to form a new enhanced feature extraction network,which uses the newly acquired 4-layer features of the improved network to detect small targets This enhanced the accuracy rate and reduced the small target pickup rate.Finally,inspired by the use of depth-separable convolution in Mobile Netv2,which greatly reduces the model parameters,the paper will use depth-separable convolution instead of the normal convolution in YOLOv4 except for the backbone feature extraction network and the convolution in the PANet structure,which contains much fewer parameters than the normal convolution,making the improved network structure smaller,less computationally intensive,and more efficient.The improved network structure is smaller,less computationally intensive,and has higher accuracy values.Due to the lack of public datasets on FOD,we create a small target foreign object dataset for airport runways,including 500 images that are similar to the actual runway road scene,and 1500 images that contain similar dimensions to the small targets to be detected in this paper are selected from the existing public dataset DOTA,and label Img software is used to label the data.