Weakly Supervised Object Detection Based On Multi-scale Feature Fusion

With the continuous development of the field of object detection,it is gradually found that the cost of labeling has become one of the important factors limiting its development.How to use simple labeled datasets to train high-precision model has become an urgent problem.To solve this problem,researchers propose a weakly supervised object detection algorithm.Although people have gradually overcome some of the challenges encountered in weakly supervised object detection,two major problems in this field have not been completely solved.The first problem is multi-scale detection.The existing detection methods are more inclined to locate and recognize the object with little scale change,so it is difficult to detect the larger and smaller objects at the same time.Multi-scale problem first appeared in the field of fully supervised detection field,but it is particularly obvious in weakly supervised detection field.The second problem is the inaccurate detection.The weakly supervised detection algorithm tends to locate the part of the object,so it is difficult to detect the complete object.In order to solve these two problems,this paper has carried out in-depth research from several aspects:(1)In order to solve the multi-scale problem,this paper proposes an improved Faster R-CNN model based on cyclic pyramid feature fusion.In this paper,through the experimental analysis,we know that the disadvantage of FPN framework is its feature fusion method.In view of this shortcoming,this paper proposes FAE module,through the top-down connection process,the feature information of each layer is added to a feature layer,so that the feature layer has stronger representation ability.In order to further fuse all the features,this paper proposes a recursive multi-scale feature fusion method,which fuses the features of each layer recursively.In this paper,the proposed framework is combined with Faster R-CNN to obtain an improved model based on cyclic pyramid feature fusion.The experimental results show that the m AP of the proposed method can reach 82.0% and 43.3% on PASCAL VOC 2007 and MS COCO datasets,which is significantly better than the existing general object detection methods.(2)In order to solve the problem of inaccurate detection,this paper proposes a weakly supervised object detection algorithm by grading the box candidates.By evaluating the quality of candidate bounding boxes generated by one-stage weakly supervised detection method,the algorithm achieves the purpose of screening candidate bounding boxes.At the same time,the algorithm also makes these candidate boxes with their confidence into pseudo labels,and trains these pseudo labels through a well-designed similar fully supervised method to get the improved model.Experimental results show that the algorithm outperforms the mainstream methods in MS COCO and PASCAL VOC datasets.(3)On the basis of the first and second parts,this paper proposes a lightweight weakly supervised object detection algorithm based on feature fusion.By replacing the second stage model of the weakly supervised detection algorithm with the YOLO which based on feature fusion,the algorithm not only solves the problem of multi-scale detection and the problem of inaccurate detection,but also achieves the purpose of improving the detection speed.At the same time,the algorithm also replaces the NMS operation in the key position with different NMS variants,which further improves the detection accuracy.Experimental results show that the proposed algorithm is superior to the mainstream weakly supervised object detection methods in both performance and speed on MS COCO dataset.