Automatic quality monitoring for steel plate is a typical challenge in industries.Although there have been several research results at home and abroad,the requirements of defect detection and recognition for low-contrast complex steel plate,are still unable to satisfy.Currently,computer vision technology based surface defect detection and recognition approaches for steel plate are increasingly concerned and taken.An effective approach can not only reduce working intensity and improve working efficiency,but also retrieve financial lost for the enterprise,which has important theoretical significance and practical application prospects.For the fact that there is less annotated data and more un-annotated data in dataset,a weakly supervised approach of surface defect location and recognition for steel plate,was presented with convolutional neural networks.First,the annotated data of defects was used to learn and extract the features by supervised convolutional neural networks.Second,the features was transferred to another net as mid-level defect representations,and then,the un-annotated data of defects was used to train the new multi-scale net.Third,by composite analysis of the multi-scale results,the final defects of the steel plate were located and recognized.Furthermore,the proposed approach has a robust and wonderful performance,which was verified by lots of experiments.