Extremal Region Analysis based Deep Learning Framework for Defects Detection

A maximal stable extremely region (MSER) analysis based convolutional neural networks (CNN) for unified detects detection framework is proposed in this paper. Our proposed framework utilizes the generality and stability of MSER to generate the desired defects candidates. Then a specific trained binary CNN classifier is adopted over the defects candidates and produces the final determined defects. Experiments on various defects dataset over different categories have demonstrated the effectiveness of our proposed framework. Beside, we can adjust the parameter settings in MSER to satisfy different requests over various industries (high precision, high recall, etc).