Research On Small Object Detection Algorithm Based On Region-based Network

Object detection is one of the key problems in computer vision tasks,and it is also the foundation for solving a series of higher-level vision tasks.It has received widespread attention and made significant progress.However,most object detection methods are dedicated to detecting objects of normal size.Small objects(less than 32*32 pixel)are usually difficult to show sufficient appearance information,and the difficulty of feature recognition and extraction is higher than that of conventional objects.However,small objects widely exist in the real world and play a vital role in many fields such as autonomous driving,defect detection,remote sensing,and smart agriculture.Therefore,how to effectively improve the feature discrimination ability for small objects,and accurately detect the location and category information of the object in the image has become the core problem in solving the small target detection task.In order to improve the detection accuracy of the small object detection algorithm,this paper proposes two improved algorithms based on the candidate region network,which are the new metrics of the two-stage detection network Faster R-CNN based on the candidate area,and improved dynamic Cascade R-CNN structure.Firstly,on the basis of two processing of the candidate region,this thesis deeply analyzes the advantages and disadvantages of IoU that was widely used in Faster R-CNN,and proposes a new measurement standard for the problem of low detection accuracy of small objects.This method abandons the characteristics of IoU series methods to calculate interaction ratio,a new metric is devised to measure bounding box similarity with truncated structurally aware distance and replace IoU in label assignment,which addresses the sensitivity of IoU to small object localization bias.On this basis,this thesis also designs a new loss function,which can describe the structural regression loss of small pests using the truncation method.Secondly,on the basis of multiple processing of the candidate region,this thesis further improves the small object detection algorithm on high-resolution remote sensing images,and proposes a multi-stage label assignment function and a multi-stage regression loss function based on a multi-stage detection network.The method of adaptively updating the threshold value is used to assign positive and negative labels in the RCNN stage,and by adaptively reducing the β threshold value,it reflects the network’s emphasis on high-quality prediction frames.The proposed method is applied to the multi-stage detection network Cascade R-CNN,and trained on the small object remote sensing dataset RSOD.For the small object detection algorithm based on the candidate area network proposed above,a large number of experiments have been carried out on the Pest24 and RSOD datasets.The detection accuracy of the method proposed in this thesis has increased by 4.7%on Pest24 and 7%on RSOD,The experimental results demonstrate the effectiveness of the method proposed in this thesis.