The Research On Robust Object Detection Algorithm With Constraints Of Scene And Object Relationship

Object detection as one of the hot research topics in the field of computer vision has a wide range of applications in daily life.In recent years,region based object detection has become the mainstream of related research.However,such methods are limited to complex real-world scenarios.In order to achieve accurate and robust object detection,this paper takes scene and object relationship context as a constraint,and studies the modeling methods of the context information in the object detection task.Based on the comparative analysis of typical algorithms,the corresponding robust object detection algorithms are proposed from two aspects,supervised detection and unsupervised detection.The main works of this paper are summarized as follows.(1)A general modeling framework for scene and object relationship context is proposed.Based on the comparative analysis of the mainstream algorithms for supervised and unsupervised object detection,a general modeling framework for scene and object relationship context for object detection tasks is summarized and proposed.Under this modeling framework,the experiments on the PASCAL VOC dataset verify the validity of the contextual relationship to the supervised object detection model,and the transferring experiments on the Cityscape to Foggy Cityscape dataset also verify the effectiveness of the contextual relationship in unsupervised detection Performance,which proves the feasibility of designing the object detection model under the proposed general modeling framework to improve the detection performance.(2)A Residual Joint Attention Network with Graph Inference(RJANet)is proposed.The proposed RJANet uses a joint attention mechanism to guide the network to selectively adjust the features from the spatial and channel dimensions,and then convert the object detection problem into a graph inference problem,and complete the reasoning under the joint constraints of scene and object relationship context to get more robust features of candidate regions.The experimental results on the PASCAL VOC dataset show that,compared with Faster R-CNN and SIN,the m AP of the proposed model is improved by 3.4% and 0.7%,respectively,proving the effectiveness of the proposed object detection model.(3)A Relation Network for Domain Adaption(RNDA)is proposed.For images in different domains,the proposed RNDA strictly aligns the local features while selectively aligning the global features by using adversarial loss.In addition,in order to introduce more task-oriented information in the unsupervised mode,RNDA uses scale dot product attention to unsupervisedly construct the required object relationship after capturing the information of the object's spatial position and visual appearance.The transferring experiments of PASCAL VOC to Watercolor and Cityscape to Foggy Cityscape show that compared with Faster R-CNN,the m AP of the proposed model is increased by 10.8% and 14.6% respectively,and compared with SWDA,the m AP of the proposed model is also increased by 0.6% and 0.5%,which proves that the proposed RNDA can reduce the difference between different domains and effectively achieve unsupervised object detection.