Weakly supervised semantic segmentation with segments and neighborhood classifiers

Semantic segmentation can provide basic semantic information for scene understanding, which has important theoretical research value and broad application prospects. Limited by the labeling cost and the scale of training data, weakly supervised semantic segmentation based on image-level labels has become a potential research issue. However, how to infer the location of image-level labels is a tough problem. Therefore, we propose a weakly-supervised semantic segmentation method with segments and neighborhood classifiers. First, we propose a scheme of segment generation based on the multiple of the number of image-level labels, which can provide high-precision boundary information with fewer regions. Second, to improve the precision of label location inference, we propose an inference method based on the most similar neighborhood granule. It can appropriately determine the number of segments contained in the inferred category label. Finally, we construct a decision table with features as conditional attribute and semantic label as decision attribute, and extract the discriminative features from attribute class reduction for neighborhood classifiers learning. Experiments evidence that our proposed algorithm can produce comparable and competitive results on widely-used MRSC and PASCAL VOC 2012 datasets.