Research On Weakly Supervised Object Detection Based On Deep Learning

In recent years,deep learning has shown exciting performance in the field of computer vision.As its important theoretical branch,object detection technology has been widely used in military and civilian fields,such as object recognition,video surveillance,satellite remote sensing and medical imaging.However,the training on strongly-supervised detection model needs to manually label the position information of all objects,which will consume a lot of time and labor costs.Therefore,weakly-supervised object detection has become an important research direction.Based on it,we carry out the research around weakly-supervised learning methods with inexact supervision.Localization and recognition are important components of object detection task.At present,weakly-supervised localization models mostly rely on the feature extraction characteristics of the network for preliminary localization.However,the hot map generated by this method tends to focus on the discriminate regions,incapable of covering entire extent of the object.To address this problem,according to the spatial dimensions and channel dimensions visualization results of different feature maps in the basic convolutional neural network,we propose a weakly-supervised object localization model based on dual-attention masks.The proposed model uses dilated convolution to retain the shallow details,and integrates the feature distribution of different dimensions with the help of attention mechanism and connected domain analysis,so as to make the network model focus on the object and reduce the interference of background information.In addition,the masking mechanism is used to force the network finding other regions of interest,improving the network's ability to extract features from the entire object area.Finally,the best structural parameters are determined through a comparative analysis of each module.The experimental results show that,compared with other methods,the proposed model in this paper has improved classification and localization accuracy,and has better ability to maintain details and grasp the overall shape of the object.Existing weakly-supervised detection models follow the two-stage detection idea based on proposal regions extraction.Since features are only extracted from the deep layers of the network,the field is relatively fixed,it is easy to occur the problems of object adhesion and poor recognition accuracy for multi-object detection.In addition,since the location and category information of the object cannot be predicted during region extraction,a large number of object proposals are needed to fit all possible objects,which will cause a lot of redundancy.To address this problem,according to the analysis results of the characteristic of different scale feature maps in the detection task,we propose a weakly supervised object detection model based on feature fusion and location filtering.The proposed model adjusts the feature distribution of the fusion feature map by fusing shallow translational variability features and deep context information to enhance its position and detail sensitivity about multi-scale objects.In addition,a new selection scheme is proposed to solve the problem of object proposals redundancy.The fusion hot map is used to generate pseudo labels instead of real labels,which effectively reduce the range of proposal regions by using the localization method.At the same time,the proposal regions are evaluated by calculating the weighting results of intersection over union and the pixel response rate for all object proposals compare with the pseudo label.Finally,the object proposals are selected with higher Io U and more tightly wrapped.The proposed model is tested on the classic detection dataset PASCAL VOC,and validating its effectiveness in terms of detection accuracy and localization accuracy.The experimental results show that the proposed model can improve the accuracy of single-class and overall classes,especially improve the detection accuracy of small-scale objects.In this paper,two important methods of class activation mapping and multiple instance learning in weakly-supervised learning are combined to optimize the model design.Based on it,we propose a new weakly-supervised object localization model and detection model.The training on proposed models only rely on image-level labels,and do not need to provide strongly-supervised boxes-level information.The experimental results show that the proposed models in this paper have a good effect on detection and localization,which is instructive for some related tasks under weakly-supervision.