| With the development of deep learning,the performance of object detection algorithms has been greatly improved,and is an important support technology for intelligent security,autonomous driving,intelligent transportation,medical diagnosis and other fields.In the practical application of object detection,there are many objects with small area share and few pixel points in the image due to factors such as angle,distance and occlusion,which have small size and limited feature information,so the detection accuracy is low and the rate of missed detection and false detection is high.In order to improve the small object detection accuracy,this paper provides an in-depth analysis of the difficulties of small object detection in different scenes and proposes two small object detection algorithms.The main work of this paper is as follows:(1)For the problem of small-scale objects miss detection arising from long-distance surveillance scenarios,a small object detection algorithm based on Residual Feature Enhanced-Symmetric Pyramid Network(RFE-SPN)is proposed.The algorithm designs a novel feature pyramid network,REF-SPN,which contains two modules,the Residual Feature Enhanced RFE and the Symmetric Pyramid Network SPN.The RFE is used to filter the background noise of the shallow feature map and enhance the underlying feature information.The SPN constructs a symmetrically structured feature pyramid network,using a two-branch structure to transfer information so that each layer can complement the high-level semantic information and the underlying detailed information from the other layers.The proposed algorithm is experimented on several data sets,and the experimental results show that the algorithm can effectively improve the small object detection accuracy and alleviate the problem of small object miss detection.(2)To address the problem of missing feature information of small objects caused by occlusion in complex scenes,a small object detection algorithm based on self-attentive networks is proposed.The algorithm proposes a Context Based Self Attention(CBSA)mechanism,which integrates contextual information mining and self-attention learning in a unified framework to model non-local features of images and effectively obtain global information.A self-attentive network CBSA-Net is designed based on the proposed self-attentive mechanism,which uses self-attentive to learn global feature information in a shallow network,strengthen the connection between distant pixels,and enhance small object features by global contextual information.Secondly,CBSA-Net reconstructs the deep feature map of the network so that the deep feature map retains as much small object information as possible without increasing the number of downsampling,while having a large perceptual field and strong semantic information.The small object detection algorithm based on self-attentive network is experimented on two datasets,and the experimental results show that the algorithm can effectively solve the problem of missing small object feature information caused by occlusion,and effectively improve the small object detection accuracy.In summary,the two small object detection algorithms proposed in this paper address the small object miss detection problem arising from long-distance surveillance scenarios and the small object information missing problem caused by occlusion in complex scenarios,and improve the small object detection accuracy by effectively using the shallow small object information as well as the global information of the network. |
