Abstract: Finding suspicious areas with algorithms of image segmentation is essential to on-line detection and recognition of surface defects for hot-rolled steel strips. Surface images of steel strips are divided into two categories：“background” and “defect”, and the two categories are modeled by hidden Markov tree(HMT) respectively and are detected at different scales. Detecting various types of defects by a same “defect” model could greatly reduce the algorithm complexity. The detection rate of common steel strip defects by hidden Markov tree model is 94.4% and the false rate is 18.8%. To reduce the steel strip’s defects segmentation false rate at finer scales, the context-adaptive hidden Markov tree model(CAHMT) is introduced and used to fuse the raw segmentations of HMT model at different scales. By the help of CAHMT, the false rate reaches 3.7%.

Key words: Hidden Markov tree model, Hot-rolled steel strip, Image segmentation, Surface inspection