Most of the solutions for weakly supervised semantic segmentation use image level supervised information to generate a class-activated feature map for training and learning. The class-activated feature map can only discover the most discriminative part of the target, which has a large gap with the real pixel-level label information, so the training effect is not satisfactory. This paper proposed an adversarial learning process on the class-activated feature map from the original image and its affine variation in order to achieve a better training effect. Firstly, it input the image and the affine image into the Siamese network to obtain their respective class-activated feature maps using the image-level classification labels, and then input the two sets of class-activated feature maps into the discrimination network for adversarial learning. The gap between feature maps and real pixel-level labels improved the performance of weak supervision. On the PASACAL VOC 2012 dataset, it achieves 63.7% mIoU scores on the validation set and 65.7% mIoU scores on the test set. Comparing with other current state-of-the-art weakly supervised semantic segmentation methods, the level-crossing ratio on the validation and test sets are improved by 1.2% and 1.3%. The proposed adversarial learning scheme can effectively utilize the equal-variable attention mechanism to learn more effective information and narrow the gap between the class-activated feature maps and the real pixel-level labels, improving the performance of weak supervision to achieve a good segmentation effect.

[2021-10-05] Printed Article /doc/2020.11.0433/v2