Weakly Supervised Object Detection Based On Deep Learning

Object detection is one of the most important topics in computer vision.The main task is to detect the existence of object categories in the picture and give positioning.Object detection has a wealth of application scenarios in medical image detection,unmanned supermarkets,and photo shopping.However,the current object detection models are full-supervised object detection models with real annotation boxes as training conditions.Frame labeling of training pictures requires a lot of manpower and material resources?In order to save the cost of labeling work,research on weakly supervised object detection came into being.The weakly supervised model only needs to label the category information of the objects in the picture,without the need to label the box information,and we can easily obtain a large number of unlabeled pictures on the Internet.Therefore,the research on weakly supervised object detection is of great significance.However,the most advanced weakly supervised object detection method is to train a weakly supervised object detection model,generate a pseudo-labeled border from the trained weakly supervised model,and train a fully supervised object detection model again using the pseudo-labeled border.This is a two-step object detection model that increases the time of object detection.This paper proposes a one-step weakly supervised model through the exploration of weakly supervised object detection,and achieves the best weakly supervised object detection effect.First,a weakly supervised object detection method based on border regression is proposed.The method is a one-step weakly supervised object detection model,and the basic multi-instance learning network is used to generate a coarse-grained pseudo-labeling frame,and the coarse-grained pseudo-labeling frame is frame-fused by the clustering algorithm to generate a fine-grained pseudo-labeling frame,and the pseudo-labeled frame is used.The weakly supervised model itself performs border regression and improves the detection results.Secondly,a weakly supervised object detection method based on graph convolutional neural network is proposed.Firstly,the model detects the frame covering the center of the object through the multi-instance network.Using the graph convolutional neural network,the object center frame and the candidate frame intersecting it are feature-fused.Let the candidate box learn the feature information of the center frame.The graph convolutional network is used to replace the redundant instance classifier optimization network,thereby improving the accuracy of weakly supervised object detection.Experiments show that the two algorithms proposed in this paper are better than the current weakly supervised object detection model in visualizing the object detection evaluation index and picture detection results.