Industrial surface defect segmentation plays a crucial strategic role in modern manufacturing and quality control, significantly impacting product quality and operational efficiency. Nevertheless, this field faces several critical challenges, including the wide spectrum of defect sizes, indistinct defect boundaries, and the complex trade-off between segmentation accuracy and computational efficiency. This paper proposes an adaptive axial squeeze transformer network (AASFormer), a novel convolutional neural network (CNN) Transformer hybrid architecture designed to address these challenges through three innovative components. First, the adaptive axial squeeze attention (AASA) mechanism dynamically adapts row and column compression rates to capture long-range contextual dependencies and fine-grained local details, enabling robust modeling of multi-scale defect geometries. Second, the edge enhancement reverse attention (EERA) mechanism synergistically integrates multi-directional Sobel operators and edge sharpening techniques with a reverse attention mechanism to sharpen defect boundaries and suppress background noise, leveraging traditional edge detection prior with deep learning attention for enhanced boundary precision, Third, the lightweight multi-scale feature fusion decoder (LMFFD) efficiently fuses hierarchical features across scales, balancing high-level semantic information with low-level spatial details to improve segmentation consistency. To facilitate comprehensive evaluation and address the scarcity of relevant datasets in this domain, a novel gear surface defect segmentation (Gear-Seg) dataset is introduced. Extensive experimental results on MSD, Gear-Seg, and NEU-Seg datasets of semantic segmentation of defects demonstrate that AASFormer consistently outperforms existing methods, achieving state-of-the-art performance in terms of mean intersection over union (MIoU) (MSD: 90.46%, Gear-Seg: 89.78%, and NEU-Seg: 84.85%).