Research And Implementation Of Small Object Detection In Complex Background

With the rapid development of convolutional neural networks in the field of image recognition,the object detection technology in general scenes has attracted much attention and has been widely used in various fields.Small object detection in complex backgrounds,as a subclass of object detection tasks,is expected to solve the problems of location and recognition of small-scale objects.However,due to the difficulties of related image acquisition,the imbalance of the ratio between objects and backgrounds,and the interference of complex background,small object detection is still confronted with huge challenges.In this paper,the research deals mainly with the problem of the poor performance of small-scale object detection in complex background.Through the analysis of visual features of small objects and the overall structure of detection networks,a region relation network based on dense pixels is proposed.In the network model,a region proposal generation method based on dense pixels is put forward to solve the problems of insufficient training samples for small-scale objects and the imbalance between positive and negative categories.The relation modeling is adopted to calculate the similarity between different regions in order to reduce the interference of the feature information in complex backgrounds when recognizing small objects.Plenty of experiments are carried out on the proposed model and an experimental comparison with other mainstream detection networks is carried out based on the SVSC Dataset--Street-View Surveillance Camera Dataset(the small-scale object detection dataset proposed in this paper).The experiment shows that region relation network based on dense pixels proposed in this paper greatly improves the accuracy when detecting small-scale objects.On the basis of this network,the detection system of surveillance cameras under street-view map is proposed.The paper studies the system's requirements and design.Next,user login management module,detection network information management module,object detection and recognition module and data management module are designed in detail.Finally,this system was implemented and tested.At present,the system has been developed and preliminary tests have been carried out.The test results show that this system can meet the needs of security personnel to detect surveillance cameras in a certain area from the street-view map,and it is an effective application of regional security analysis in the urban scene understanding.