Box-driven Weakly Supervised Images Semantic Segmentation Algorithm Based on Attention Model
Jianxin Liu, Yushui Geng, Jing Zhao, Wenxiao Li, Kang Zhang · 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC) · 2021
Image semantic segmentation has made great progress using deep neural networks. However, methods based on deep neural networks rely on numerous pixel-level annotations, and such pixel-level annotations require considerable time and money. To overcome this problem, researchers have proposed a weakly supervised image semantic segmentation, which requires only simple annotations to achieve high segmentation accuracy. In this article, we propose an algorithm that uses bounding box annotations to achieve image semantic segmentation. The basic steps are divided into two steps: generating temporary semantic segmentation labels and training the network and then iterating these two steps. We have conducted many experiments on PASCAL VOC 2012 to prove that this algorithm is effective, and the performance improved, which is close to the performance of the corresponding fully supervised network.