Abstract: A method for detecting a surface defect based on a perceptual autoencoder is provided, which belongs to the field of integrated circuit defect detection. By using perceptual loss instead of pixel-by-pixel loss, the problem of blurred single-pixel-loss output image with serious distortion at the edges is solved. This manner is highly migratory, and it may migrate well to the new target to be detected. On the other hand, by using the feature pyramid instead of the sliding window manner, the feature pyramid does not require a plurality of redundant computation by sliding on the original image, which effectively reduces the memory and computation cost. A system for detecting a surface defect based on a perceptual autoencoder is also provided, which includes specific modules for implementing the above method. In addition, a computer device and a storage medium of the above method are provided, for accomplishing operations of the above method.