Weakly supervised object detection methods have achieved great success in natural scenes. However, they confront challenges such as small and weak appearances when detecting surface defects in industrial scenes. To address this problem, we propose a global-local Context Mutual Evolution Network (CMENet) for weakly supervised surface defect detection. CMENet is a dual-branch network consisting of a CNN branch and a Transformer branch, in which a context mutual evolution (CME) module is introduced in multiple blocks to enhance the feature representation ability of the network through the global and local information interaction. The CME module helps the network focus on defective areas and suppress background interference via a Feature Cluster Attention (FCA) module. The FCA adaptively selects some meaningful tokens as keys and values by a learnable clustering module to calculate attention. Extensive experiments on the DAGM 2007, KolektorSDD2, and Magnetic Tile defect datasets demonstrate that our method achieves promising performance compared with other state-of-the-art methods.