Distilled fine-grained domain adversarial network prompted by normalization and regularization

Abstract

Fine-grained domain adaptation presents a significant challenge in machine learning, where subtle differences between classes and domains often lead to poor generalization. Addressing this issue is crucial for improving the accuracy and robustness of models when applied to unseen data from different but related domains. We propose a distilled fine-grained domain adversarial network in which feature distributions across diverse domains are effectively aligned, and classification discrimination is enhanced during the mini-batch training phase. Based on knowledge distillation, the capability for fine-grained feature extraction is transferred from the teacher network to the student feature extractor, all while preserving the original data and prediction distributions. Consequently, low-entropy prediction distributions can be effectively leveraged by the domain adversarial network for unsupervised training in the target domain, leading to notable improvements in fine-grained recognition performance. In order to train the domain adversarial network more effectively, a linear transformation bottleneck layer for feature normalization is introduced after the feature extractor. This layer helps alleviate the normalization conflict between training and inference, which arises due to domain shift in mini-batches of data. Moreover, to strike a balance between the competing effects of the classifier and discriminator, we devise a double bilinear dropout module in the domain discriminator.