Optimal Scale of Hierarchical Image Segmentation with Scribbles Guidance for Weakly Supervised Semantic Segmentation

作     者：Al-Huda, Zaid Zhai, Donghai Yang, Yan Algburi, Riyadh Nazar Ali 

作者机构：Southwest Jiaotong Univ Sch Comp & Artificial Intelligent Chengdu 610031 Sichuan Peoples R China Southwest Jiaotong Univ Sch Mech Engn Chengdu 610031 Sichuan Peoples R China 

出 版 物：《INTERNATIONAL JOURNAL OF PATTERN RECOGNITION AND ARTIFICIAL INTELLIGENCE》 (国际图形识别与人工智能杂志)

年 卷 期：2021年第35卷第10期

核心收录：

学科分类：08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：National Science Foundation of China [61772435  61976247  61961038] 

主　　题：Deep convolutional neural networks graphical model hierarchical image segmentation semantic segmentation weakly supervised segmentation 

摘      要：Deep convolutional neural networks (DCNNs) trained on the pixel-level annotated images have achieved improvements in semantic segmentation. Due to the high cost of labeling training data, their applications may have great limitation. However, weakly supervised segmentation approaches can significantly reduce human labeling efforts. In this paper, we introduce a new framework to generate high-quality initial pixel-level annotations. By using a hierarchical image segmentation algorithm to predict the boundary map, we select the optimal scale of high-quality hierarchies. In the initialization step, scribble annotations and the saliency map are combined to construct a graphic model over the optimal scale segmentation. By solving the minimal cut problem, it can spread information from scribbles to unmarked regions. In the training process, the segmentation network is trained by using the initial pixel-level annotations. To iteratively optimize the segmentation, we use a graphical model to refine segmentation masks and retrain the segmentation network to get more precise pixel-level annotations. The experimental results on Pascal VOC 2012 dataset demonstrate that the proposed framework outperforms most of weakly supervised semantic segmentation methods and achieves the state-of-the-art performance, which is 69.3% mIoU.