Defects on steel strip surface can long-term cause undesirable effects, since they make physical and/or chemical properties mismatched from steel's specification. Nowadays, automatic visual-based surface inspection is adopted, in order to detect the defects on steel strip surface after being produced. Moreover, since these defects appear in wide variety of forms and various classes, machine learning methods are generally involved to visual surface inspection for coping with these appearances. In this paper, we present a novel defect detection model to perform defect segmentation of hot-rolled steel strip surface, by using Convolutional Auto-Encoder (CAE) and sharpening process to extract the defect features of input image, then applied postprocessing for visualization. In the experiments, the NEU database, which provides six kinds of typical surface defects of hot-rolled steel strip, was applied to evaluate the efficiency of the proposed model. This database also provides difficulty challenges regarding diversity of intra-class and similarity of inter-class. The results show that the proposed model can perform defect segmentation in all kinds of defects in database, however the efficiency was compromised by illumination changes. Notable that, this segmentation is based on unsupervised learning with small training dataset and no labeling procedure, so it can be easily extended to the real world application. Eventually, this defect detection shall improve the productivity and reliability of steel strip's production process.