Unsupervised Salient Object Detection with Momentum Expansion: A Case Study on the Xiongse Monastery in Xizang
Qian Luo, Xiaobo Yang, Jianbang Li, Yincheng Xu · Research Square · 2024
Abstract A distinctive and intricate color palette is often used for interior decorations in Xizang temples. Traditional methods of detection can be challenging, as they are prone to being limited by similar background colors, making it difficult to identify all the essential elements. In this case, a momentum-based pseudo-labeling strategy is proposed to address this issue. This strategy utilizes momentum expansion to categorize the pseudo-labels generated by the first-stage pre-trained model into active areas, expansion areas, and background areas, they facilitating the supervision of fine-label generation in the second stage. Additionally, a real-time label correction strategy is employed. An Efficient Channel Attention (ECA) module is introduced to enhance the model's feature learning capabilities. An unsupervised deep learning-based object detection method is utilized, eliminating the need for extensive manual annotation of labels for diverse objects. Experimental results have demonstrated the significant effectiveness of this approach in comprehensive detection of salient region targets.