Research On Small Object Detection Algorithms In Complex Scene

In recent years,object detection,which has developed rapidly,has been integrated into all aspects of people's lives.With several years of development,a lot of object detection algorithms have been proposed.The accuracy of object detection is getting higher and higher.However,the effect of small object detection is not satisfactory.The features extracted from small object are few since the small object is too small,which lead to the difficulty of small object detection.Small object detection requires multiple scale feature.The current object detection uses the backbone network combined with feature pyramid networks(FPN)to extract different scale features.FPN only fuses the features of adjacent layers during feature fusion.The features of non-adjacent layers are diluted,which is unfavorable for small object detection.The current object detection algorithms get region of interest from the manual anchor,which has a single shape and cannot cover small targets well,resulting in a low object recall rate.The manual anchor is not suitable for small object detection.This paper proposes a Dense fused feature pyramid network(Dense FPN)based on above shortcoming.Dense FPN connects the feature maps of each scale to the feature pyramid by skip connection.The feature pyramid can directly fuse all the scale feature map,which reduces the consumption of information in the network transmission process and increases the richness of each feature of the feature pyramid.In addition,this paper proposes a Deformable Convolutional Network Based Guided Anchoring(DCN-GA)based on deformable convolution.DCN-GA generates anchor boxes of various shapes and sizes through the network without preset anchors.The DCNGA uses deformable convolution to generate various shapes anchor to match small size irregularities.In the end of this paper,Dense FPN and DCN-GA are integrated in one network.This paper experiments on the MS COCO dataset and the automatic driving traffic signs based on virtual simulation environment dataset,and compares the experimental results with the existing small object detection algorithms.The experimental results show that the proposed method effectively improves the accuracy of small object detection.