Disentangling the intrinsic feature from the related feature in image classification using knowledge distillation and object replacement

Abstract

In our previous work, the features extracted by deep learning models are conceptually divided into the intrinsic features and the related features, where the object-oriented features are the intrinsic features and the background-oriented features are the related features. And the intrinsic feature-based classification results are regarded more reliable. In this work, we have improved our previously used metric and propose a new metric called Reliable Classification Accuracy (RCA) to more effectively quantify the dependency of classification results on the intrinsic features. Compared to our previous method that heightens the intrinsic feature exaction, we propose a novel method to further disentangle the intrinsic features from the related features. The proposed method comprises specially designed knowledge distillation and unique data augmentation. Unlike traditional knowledge distillation, the teacher model shares an identical network structure with the student model but has a special training process: the teacher model is pre-trained on a carefully designed dataset for pure intrinsic feature extraction and then guides the student model to focus on the category-related objects while ignoring the backgrounds in general datasets. And in the unique data augmentation, the category-related objects and the backgrounds from different images are re-matched in a graph-based method to further disentangle the intrinsic features from the related features in the guided feature extraction by specially designed knowledge distillation. The experiment results prove that the proposed method significantly improves both RCA and classification accuracy on the filtered datasets from MS COCO, Pascal VOC and Open Image Dataset.