Research On Image Classification Based On Contrastive Self-learning

Stable semantic features are a prerequisite for achieving exceptional image classification,but traditional methods of semantic supervision through manual labeling suffer from high labor costs and inadequate generalization of semantic labels.To address these problems,contrastive learning is introduced to achieve feature self-learning.Hence,two implementation solutions are provided for self-supervised image classification in the fields of deep image clustering and partial label learning.In the deep image clustering task,a pseudo-supervised clustering method based on contrastive selflearning is proposed,which improves the clustering performance by assigning pseudolabels globally.In the partial label image disambiguation task,a contrastive selflearning-based partial label disambiguation framework is designed,which improves the semantic representation and effectively alleviates the dual uncertainty in partial label learning,thereby improving the performance of partial label image disambiguation.These methods obtain relatively stable semantics by mining the transformation invariance of images,which can achieve better image classification results as well as avoid the problems of high cost and poor generalization ability of manual labeling.This provides a feasible path for model self-learning and promotes the development of intelligent agents from "artificial" intelligence to "artificial intelligence".The research contents and main results of this thesis include:(1)To solve the problem of degraded clustering performance due to unstable feature semantics in deep image clustering,a pseudo-supervised clustering method based on contrastive self-learning is proposed.Firstly,the importance of stable semantics for image classification performance is clarified,and stable semantic features(i.e.,metafeatures)are abstracted based on the instance-level semantic features obtained from contrastive learning.Then,a global pseudo-label assignment mechanism based on metafeatures is proposed for the first time,which effectively reduces the misclassification problem of feature semantics at the classification boundary.We introduce the labelsmoothing cross-entropy loss,which not only avoids the overconfident prediction of pseudo-labels but also directly maps features to semantic labels.Experimental results show that this method extracts stable semantic features without the need for manual labeling and achieves remarkable clustering performance,outperforming existing clustering methods.(2)To solve the dual uncertainty problem between representation learning and partial label disambiguation methods in image partial label disambiguation,a contrastive self-learning based on the partial label disambiguation method is proposed.First,semantic-aware data augmentation is introduced in contrastive learning to enhance the semantic expression ability of features.Then,stable semantic features(i.e.,metafeatures)corresponding to the relevant categories are selected from the embedding features for contrastive disambiguation,improving the accuracy and robustness of the model.Finally,we jointly train the contrastive representation learning module and the meta-feature contrastive disambiguation module,which can effectively alleviate the dual uncertainty problem between representation learning and disambiguation methods.Experimental results show that this method improves the performance of partial label disambiguation without the need for accurate manual labeling and outperforms existing partial label learning methods.In summary,this thesis proposes two self-supervised image classification methods based on contrastive self-learning,which combine contrastive learning,deep clustering methods,and partial label disambiguation methods from the perspective of stable semantics.The dependence on image classification models is avoided on a large number of manually labeled data and self-learning is achieved in two classification tasks.The experimental results demonstrate the effectiveness and superiority of the proposed methods and provide a new way for self-learning of models in image classification tasks.