Research On Object Detection Algorithm Based On Weakly Supervised Learning

Object detection is an important task in the field of computer vision,which requires accurately detecting objects in images and providing their location information.Traditional object detection models require a large amount of annotated data,including object category and bounding box annotations,which require a significant amount of human and time costs.In contrast,weakly supervised object detection only requires object category information to learn from unannotated image data,thus greatly reducing annotation costs.In addition,weakly supervised object detection can be applied to a wider range of scenarios,such as medical image detection,supermarket product recognition,and autonomous driving.Therefore,studying weakly supervised object detection is of great significance for reducing annotation costs and expanding application scenarios.This thesis aims to study weakly supervised learning-based object detection algorithms to improve the performance of weakly supervised object detection and expand its application range.The main contributions of this thesis are:A weakly supervised object detection algorithm incorporating pyramid squeeze attention and candidate box selection is proposed.The algorithm improves upon the Online Instance Classifier Refinement(OICR)model,which is based on weakly supervised learning for object detection.Firstly,a Pyramid Squeeze Attention(PSA)module is added to the backbone feature extraction network to effectively extract implicit positional information from the features,enhancing the network’s localization ability and enabling better consideration of the global features of the targets.Secondly,a Candidate Box Screening Algorithm(CBSA)is introduced for candidate box selection.This algorithm selects the best candidate boxes based on confidence scores,sizes,positions,and other information provided by the network.These selected candidate boxes are used as pseudo ground-truth boxes for the regression module,thereby improving detection accuracy.A weakly supervised object detection algorithm based on candidate box generation is proposed.The algorithm improves upon the Online Instance Classifier Refinement(OICR)model,which is based on weakly supervised learning for object detection.Regarding candidate box generation,the Selective Search algorithm is modified and combined with Ablation-CAM.This new candidate box generation algorithm produces more accurate candidate boxes that better encapsulate the entire object,resulting in higher intersection-over-union with the ground truth bounding boxes.By improving the generation of candidate boxes,the algorithm enhances detection accuracy and effectively improves the performance of weakly supervised object detection.Finally,both approaches are combined into a single model,resulting in a weakly supervised object detection algorithm based on candidate box generation and screening.Experimental results on the PASCAL VOC 2007 dataset demonstrate that the proposed weakly supervised object detection method improves detection accuracy and localization accuracy to a certain extent.