| With the development of deep learning,fully supervised learning has made great progress in many fields such as computer vision,natural language processing,and speech recognition.Fully supervised learning relies on a large amount of data annotation,but large-scale refined annotation requires a high cost,which affects the further development of deep learning.Weakly supervised learning only needs incomplete,inaccurate or inexact data annotations to complete the same machine learning tasks as fully supervised learning.Therefore,research on weakly supervised learning has important practical application significance.This thesis focuses on the research of object localization with image-level category labels as weakly supervised labels.For solving the problems encountered in the existing weakly supervised object localization methods such as the inability to mine the entire area of the object,the locating ability of the unmined intermediate convolutional layer,and the blurring of the object edge in the localization map.The main research results obtained include:(1)A Gradient-based method of Refined the Class Activation Map is proposed.Aiming at the problem that the existing methods only locate the object local area,by extracting the correlation information between the probability values of each class output by the classification network,and integrating this correlation information into the feature map to enhance the information of a specific category.The feature map has stronger weak supervision and characterization capabilities,thereby enhancing the integrity of object localization and improving localization accuracy.Experiments on the ILSVRC classification data set show that the proposed method is significantly better than the existing weakly supervised localization methods,and the Top-1 loc error reaches 56.48%.(2)A Dual-Gradients Localization framework for weakly supervised object localization is proposed.Although the Gradient-based method of Refined the Class Activation Map method expands the localization area,the localization capability of the middle convolutional layer has not been tapped.In response to this problem,by introducing the gradient of a specific category on the intermediate convolutional layer and the gradient of cross entropy on the intermediate convolutional layer,to enhance and extract the information about the specific category in the features map of the intermediate convolutional layer,so as to improve the characteristics of the intermediate convolutional layer.Experiments on the ILSVRC dataset show that the middle convolutional layer of the classification model has localization capabilities,and the Top-1 loc error reaches52.23%.(3)A Dual-Gradients Localization framework with Skip-Connection for weakly supervised object localization is proposed.Although the above two algorithms expand the localization area and mine the localization ability of the intermediate convolutional layer,the edges of the localization map are not accurate due to the lack of use of edge information.To solve this problem,in the Dual-Gradients Localization framework,shallow features are introduced to increase the edge information of the object,so as to achieve more accurate localization.Experiments on the ILSVRC data set show that the convolutional feature map that introduces edge information has better localization and characterization capabilities,and the Top-1 loc error reaches 43.14%. |
