Research On Weakly Supervised Object Localization And Detection Based On Class Activation Map

At present,fully supervised learning has been widely used in speech recognition,natural language processing,computer vision and other fields.However,fully supervised learning heavily relies on large-scale annotated datasets,which often cost lot of labor,material resources and time.Weakly supervised learning only use incomplete,inaccurate or inexact labeled data to complete the same tasks as fully supervised learning,and can save labor and time costs.Therefore,weakly supervised learning research is of great significance for the development of machine learning.This thesis studies the incomplete object detection in weakly supervised object detection via deep learning,and the main work is as follows:(1)This thesis has proposed a spatial divergence and feature map accumulation(SD-FMA)method for weakly supervised object localization.This work fuses the edge and contour information of the objects in the shallow layer into deep features by fusing the feature layers of the convolutional neural network,which can enlarge the activation area.At the same time,this thesis has proposed a spatial difference loss function to force different feature maps be activated as much as possible.Finally,the feature map accumulation method is used to expand the detection area of the object.Experimental result shows this method can achieve good weakly supervised object localization.(2)This thesis has proposed a regional similarity module network(RSMNet)for weakly supervised object localization.Due to in view of the similarity in the semantics of objects themselves,a similarity module is firstly proposed to calculate the features similarity between any two pixels to obtain a similarity matrix.Then,RSMNet multiplies the similarity matrix and the class activation map in the classification network to achieve the similar activation map(SAM).Meanwhile,the background regularization loss and difference regularization loss are proposed to suppress the background activation.Finally,this thesis adopts a feature map accumulation method to fuse the feature maps of multiple channels and obtain a complete activation map.(3)In order to address the problem of weakly supervised object detection,this thesis adopt weakly supervised object localization to further develop a weakly supervised object detection method.This work first uses conditional random field(CRF)to refine the activation map generated in RSMNet to obtain pseudo masks.Then,this work adopts a semantic segmentation network via these pseudo-masks to achieve the pixel-level masks.Using the pixel-level masks can get the pseudo bounding box which are used to train the object detection network.In this way,weakly supervised object detection is achieved.Since the poor quality of the pseudomask may seriously affect the performance of the network,this work adopts a cyclic training strategy to train the segmentation network,and replaces the poor-quality pseudo-mask by the mask output by the network,which can improve the performance of the segmentation network.