A novel anomaly detection method named EDA (enhancing anomaly detection via adversarial anomaly learning) was proposed, to address the challenges of industrial image anomaly detection, including the scarcity of anomalous samples, the complexity of annotation, and the high computational cost of deep models. The proposed approach consisted of two key stages. 1) Anomaly learning and embedding stage: a generative adversarial network (GAN) architecture was employed to learn anomalous features. The generator’s parameters were reduced to ensure lightweight design, and subpixel convolution was introduced to enhance anomalous information. Random regions were selected from normal images, refined using the SAM (segment anything) model, and then anomalous features were generated in these refined regions, providing prior anomalous features and corresponding masks for the anomaly detection stage. 2) Anomaly detection stage: a Contrast U-net network was introduced to improve sensitivity to anomalous features and enhance the accuracy of identification and localization through supervised training. Experimental results on the MVTec dataset demonstrated the superior performance of the proposed method, achieving an image-level AUROC of 98.2%, a pixel-level AUROC of 97.8%, and an AU-PR of 81.1%, showing significant advantages and outstanding performance in the field of industrial image anomaly detection and segmentation.