Defect Analysis of Inner-Wall of Pipes by Differentiated Residual Blocks of Convolutional Neural Network

Vision-based approaches are increasingly applied in automatic systems of defect analysis to solve the problems of manual assessment such as subjective, time-consuming, tediousness and high cost when using the laser, X-ray, etc. This kind of issues targeted in the research is inside-pipe defects, which is quite new and difficult for traditional Image Process techniques because of complex illumination, background and defect textures. To deal with these problems, we propose a method using Aggregated Residual Network, one of the top current state-of-the-art Convolutional Neural Networks, which is able to overcome most weakness of ones using hand-craft features in arbitrary conditions and a multi-scale detection approach which can deal with the drawback of class imbalance and the variety of defects like shapes, sizes, contours, etc. coming from the practical data set collected in industrial environment. Empirical results show the efficiency of the proposed methodology, where the accuracy of defect recognition is approximately human-like, over 99.5%, and defect detection is remarkable, over 97.3%.