Multi-granularity Sequential Three-way Decision Method For Object Detection Based On Deep Learning

Object detection methods based on deep learning are important research directions in computer visual,they are widely used in real-world scenarios,including face detection,autonomous vehicle and medical diagnosis.Traditional object detection attempt to achieve a high precision,regarding the misclassification costs as the same and it is unreasonable in practical scenarios.The objects have different importance,all kinds of misclassification will lead to different costs.It is the imbalanced cost issue.And the object detection task need sufficient information,it is a multi-stage decision process,the effective decision need a reasonable range of information.It is the insufficient information issue.The imbalanced cost issue and the insufficient information issue are exist in object detection application,they will bring challenges to traditional methods.To address these issues,we propose a multi-granularity sequential three-way decision(MGS3WD)method for object detection based on deep learning,design optimal decision with granular computing,sequential three-way decision and cost sensitive learning.Firstly,we analyze the object detection method based on deep learning,discuss the architecture of deep convolution neural network and region proposal network,design the method of extract the feature map from object.Secondly,we regard the object detection as a multi-stage decision process,considering the process of object information from rough granularity to precise granularity as sequential decision in granular computing perspective,introducing the delayed decision in three-way decision to address the insufficient information issue.Thirdly,we denote the costs in decision with costsensitive learning,design the optimal method to minimize the total cost,to address the imbalanced cost issue.Then,we conduct experiments on Pascal and COCO object detection database,using VGGNet and Res Net object feature extract network in Faster R-CNN method,comparing the proposed method with traditional two-way decision in cost,error rate and running time.Finally,we evaluate the proposed method with object classes,balance parameter,sample number,cost setting and detection threshold,these experiments demonstrate the proposed method has better perform than traditional twoway decision with experimental setting,it can address the imbalanced cost issue and insufficient information issue effectively.