This paper is concerned with the problem of automatic inspection of hot-rolled plate surface using computer vision. An automated visual inspection system has been developed to take images of external hot-rolled plate surfaces and an intelligent surface defect detection paradigm based on gradient spectrum technique is presented. Gradient spectrum characterizes the spatial configuration of local image texture and is robust against any monotonic transformation of the gray scale. Texture features based on gradient spectrum are extracted from ROI in hot-rolled plate surface images and integrated into a feature vector which uniquely differentiates the abnormal regions from normal surface. Classification accuracies using the gradient spectrum and gradient-based method are compared. The results indicate that gradient spectrum performs well in classifying the samples with the lowest classification error.