WEAKLY SUPERVISED SEMANTIC SEGMENTATION BASED ON MULTI-LEVEL ONLINE CLASS ACTIVATION MAPPING LEARNING

Aiming at the problems of incomplete class activation maps and detailed information loss in weakly supervised semantic segmentation task, this paper proposes a multi-level online accumulation framework for class activation maps learning. It included a refined gradient class activation mapping generation module and a multilayer online accumulation of class activation mappings module. The refined gradient class activation mapping generation module used the gradient information to generate more accurate class activation map. Multilayer online accumulation of class activation mappings module iteratively fused the class activation maps which were generated by each layer. This module combined the result of online class activation map in each layer to obtain the final class activation map. Experiments adapted PASCAL VOC2012 as dataset. Experiments show that this algorithm has good segmentation effect and greatly improves the segmentation accuracy, which verify the effectiveness of the proposed algorithm.