One-Shot Municipal Solid Waste Detection via Object-Relevant Feature Enhancement and Category-Level Feature Fusion

Due to the significant variation of municipal solid wastes in appearance and composition, as well as the lack of abundant labeled samples, deep learning-based municipal solid waste detection is a challenging problem. This paper presents a novel one-shot municipal solid waste detection model based on Faster R-CNN and the attention mechanism to improve detection performance via object-relevant feature enhancement and category-level feature fusion. Concretely, a spatial attention-based feature enhancement module, SAFEM, is designed to enhance object-relevant information and improve object localization. Then, the channel attention-based fusion module, CAFM, is proposed and applied in two stages separately. In the first stage, CAFM uses the category-level information of the support features to help the region proposal network filter out non-support category query proposals; in the second stage, CAFM is used to enhance the classification accuracy of support category objects. The effectiveness of the proposed model is verified by experiments on the waste dataset of Huawei Cloud Competition 2020. The experimental results demonstrate that the proposed model achieves remarkable performance in one-shot municipal solid waste detection.