A unified framework for interactive image segmentation via Fisher rules

Interactive image segmentation has been an active research topic in image processing and computer graphics. One of its appealing advantage is the optimization of human feedback and interactions to generate user-desired results. Segmentation results of most previous methods usually depend on the quality and quantity of the user input. In this paper, we propose an algorithm to solve one important challenge arising from inputs with bad or limited quality. Our work is notably different from previous methods. First, the weakly interactive image segmentation is formulated and deduced in theory, then we propose to reconstruct enough samples via sparse reconstruction to enhance the robustness to weakly interactive labels. More importantly, we leverage interactive labels to extract a latent subspace which jointly optimizes multiclass classification and binary classification based on fisher rules. Numerous experiments are conducted on MSRC (Ning et al. in Interact Imaging Vis Pattern Recognit 43(2):445–456, 2010) and KIM (Kim et al. in: 2010 IEEE computer society conference on computer vision and pattern recognition, 2010) database. The results demonstrate effectiveness and efficiency of our method.