Weakly Supervised Semantic Segmentation (WSSS) aims to perform pixel-level semantic segmentation using only image-level annotations. However, the sparse supervision and frequent co-occurrence of foreground and background regions make this task challenging. Most existing approaches rely on Class Activation Maps (CAMs) to generate pseudo-labels, which often suffer from false activations and limited accuracy, constraining further performance gains. To overcome these limitations, this paper proposes a novel framework, Adaptive Prototype Learning for Weakly Supervised Semantic Segmentation (CPDGL). Central to this approach is the Prototype Knowledge Refinement Module (PKRM), which constructs class prototypes from class tokens extracted by a Vision Transformer (ViT) and employs an Exponential Moving Average (EMA) mechanism to ensure stable and progressive prototype updates. Building upon this, the proposed Prototype Contrastive Loss ( \(\mathcal {L}_{PCL}\) ) encourages features to align closely with their corresponding class prototypes while enhancing inter-class separability, thereby improving category discriminability and semantic cohesion. To further strengthen local feature learning, the Local-Global Discriminative Fusion Module (LGDFM) is introduced, integrating fine-grained local semantics with global class-level information. Concurrently, a Local Consistency Contrastive Loss ( \(\mathcal {L}_{LCL}\) ) is introduced to suppress pseudo-label noise and improve the semantic consistency of local feature representations. Experimental results show that CPDGL achieves 72.5% and 71.8% MIoU on the PASCAL VOC validation and test sets, and 42.5% on the MS COCO validation set, outperforming mainstream one-stage methods and demonstrating its effectiveness and generalization in WSSS.