Partially Occluded Self-taught Learning via Residual Connection Union
Bin Kang, Jianchun Ye, Сонглин Ду, Xin Li · 2022 15th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI) · 2022
Self-taught learning is a promising technology that utilizes easily accessible label-free images to narrow the performance gap in few-shot learning. During data collection, it naturally comes along with numerous partially occluded samples that miss discriminative information. However, few works have seriously studied the impact of partially occluded samples on self-taught learning. In this paper, we propose a cross combination oriented auto-encoder for partially occluded self-taught learning. The proposed auto-encoder is composed of multi-scale Residual Connection Unions (RCU). The innovation of RCU lies in its capability of integrating local attention and self-attention into a united structure, which facilitates exploring the potential relation between convolution and self-attention operators under a different receptive field. Extensive experiments on our built and Caltech 101 datasets demonstrate that the proposed method sharply outperforms the existing state-of-the-art competitors in heavily occluded cases.