Feature-embedding-based memory bank methods have proven effective in the task of image anomaly detection. However, they face challenges due to domain bias from pre-trained networks, insufficient representativeness of the memory bank, and feature space bias, which ultimately degrade detection performance. To address these issues, this study proposes a bias-reduced coupled-hypersphere model for image anomaly detection. First, an anomaly generation method to create a self-supervised learning task for model fine-tuning is proposed, which mitigates domain bias. Second, core-set sampling is employed to establish a more representative memory bank of normal features. Then, coupled hyperspheres are identified within the memory bank to differentiate between normal and abnormal features through model training. During the inference stage, local density K-nearest neighbor (KNN) is applied to minimize the effects of feature space bias in the memory bank. Experimental results on the public MVTec AD industrial dataset show that the proposed method achieves image- and pixel-level area under the receiver operating characteristic curve (AUROC) scores of 99.5% and 98.4%, respectively, outperforming most existing approaches.