Research On Class Activation MAP Generation Based On Class Semantic Relations

As an important carrier of information carrying and dissemination,images are one of the important research objects in the field of deep learning.With the development of information technology and the arrival of the era of big data,the amount of image data is exploding.For image processing tasks in deep learning,a large amount of image data requires a large amount of labeling information.Manually labeling strong supervised labels is very time-consuming and labor-intensive,and the image database used in practical applications often lacks sufficient strong supervised labels.Because the weakly supervised labels have the characteristics of simple labeling,the researchers turned to the research of weakly supervised image processing tasks.Compared with strong supervised image processing,weakly supervised image processing tasks have the advantages of low label labeling cost and rich training data.Therefore,research on weakly supervised image processing tasks has a positive significance for breaking through the bottleneck that current strongly supervised image processing tasks rely too much on training data.However,the image-level weak supervision label only labels the category information contained in the image,and faces the challenge of converting the image-level category information into pixel-level object region information.In recent years,researchers have proposed a class-based activation map extraction method based on deep learning.Since the discriminable information of objects can be captured in the deep semantic space,the class activation map method effectively improves the performance of object localization.However,the existing category activation map methods focus on a single classification model that considers all categories,and the discriminable information provided is limited,resulting in the extracted regions being local regions rather than global regions.To this end,this thesis builds multiple classification networks with diverse discriminable information based on class semantic relations,and considers constructing a new discriminative map extraction network model based on orthogonality,thereby substantially improving the performance of activation map extraction.details as follows:1.This thesis proposes a class activation map extraction method based on class selection.The traditional class activation map extraction method uses all categories in the data set to train a single network without considering the relationships between the categories.The extracted discriminative information is limited,resulting in the activation map that is usually incomplete and only a locally discriminable region.Aiming at this problem,this thesis studies the relationship between the categories and introduces the category selection method based on the differences between the categories.The category pairs are constructed by selecting the representative categories,and multiple binary classification networks are trained to extract the class activation map.Aiming at the problem of low-level feature information loss,this thesis proposes a different level feature maps fusion structure.Through the fusion of high-level and low-level feature information,the performance of class activation map extraction results is effectively improved.2.This thesis proposes a class activation map extraction method based on class clustering.Because there are too many models to be trained in the method of extracting activation map based on category selection,it is more complicated.Aiming at this problem,this thesis further studies the relationship between the categories,starting from the similarity between the categories,by clustering similar categories into category clusters as a new category training classification model,and training a small number of multi-class networks to extract the class activation map.The class activation map extracted by this method are complementary.By merging multiple class activation map,good performance is obtained.3.This thesis proposes a class activation map extraction method based on feature orthogonality.The methods proposed in this thesis based on category selection and category clustering-based activation map extraction are based on category relationships.There is no need to improve the classification model for the task of map enhancement.Therefore,this thesis introduces a feature orthogonal module and a two-branch classification network.By forcing the features to be orthogonal between the two sub-networks of the two-branch classification network,the discriminative ability of deep learning features is enhanced.Finally,the two-branch classification network with feature orthogonal module and class clustering method are combined to further improve the performance of class activation map extraction.