Though image-level weakly supervised semantic segmentation (WSSS) has achieved great progress with Class Activation Map (CAM) as the cornerstone, the large supervision gap between classification and segmentation still hampers the model to generate more complete and precise pseudo masks