A Weakly-Supervised Approach for Semantic Segmentation
Yanqing Feng, Lunwen Wang · 2019 IEEE 3rd Information Technology, Networking, Electronic and Automation Control Conference (ITNEC) · 2019
Deep learning shows great power in many computer vision tasks, however these methods require a massive labeled samples to train the neural networks. In the task of semantic segmentation, it is usually unrealistic to obtain such a large number of labeled training samples. The weakly supervised method provides a promising idea to solve this issue using weak supervised information such as the image level labels. In this paper, we propose a weakly supervised approach for semantic segmentation. In the proposed approach, the orthogonal non-negative matrix factorization (ONMF) is utilized to extract the features of different targets with strict sparse constraints. The gradient optimization is used to solve the objective equation of ONMF. To start the iteration with a preferable initial value, we adopt the mini-batch k-means approach to cluster the massive training samples with image level labels and use the clustering centers as the initial value of the base matrix, which accelerates the convergence of the iteration effectively. using the image label in the reconstruction of the image, select the features corresponding to the image label, then the probability of every pixel can be obtained. The effectiveness of the proposed approach is demonstrated through the open dataset.