<p>Anomaly detection (AD) has been extensively studied and applied across various scenarios in recent years. However, gaps remain between the current performance and the desired recognition accuracy required for practical applications. This paper analyzes two fundamental failure cases in the baseline AD model and identifies key reasons that limit the recognition accuracy of existing approaches. Specifically, by Case-1, we found that the main reason detrimental to current AD methods is that the inputs to the recovery model contain a large number of detailed features to be recovered, which leads to the normal/abnormal area has not/has been recovered into its original state. By Case-2, we surprisingly found that the abnormal area that cannot be recognized in image-level representations can be easily recognized in the feature-level representation. Based on the above observations, we propose a novel recover-then-discriminate (ReDi) framework for AD. ReDi takes a self-generated feature map (e.g., histogram of oriented gradients) and a selected prompted image as explicit input information to address the identified in Case-1. Additionally, a feature-level discriminative network is introduced to amplify abnormal differences between the recovered and input representations. Extensive experiments on two widely used yet challenging AD datasets demonstrate that ReDi achieves state-of-the-art recognition accuracy.</p>