Statistics Based Neural Networks Method for Industrial Image Inspection

Automated industrial image inspection system has attracted a great deal of interest in recent years. In this thesis, a new method is presented by combining a statistics method with a neural networks method, which could reduce the interference of machine dynamics and improve the inspection accuracy. Different from the pixel-based or feature-based methods, the proposed method is based on two indices of an image, which are the variances of the rows and columns of the image. For image inspection, first neural networks are trained using these two indices from a set of good images in order to establish a tolerance zone. Then, the two indices of each inspection image are computed through trained neural networks and compared with the tolerance zone. A defective item is detected if either index falls out of the tolerance zone. The other contributions, such as two-point based image registration method and defect simulation algorithms, also help to improve the performance of inspection. Experimental results demonstrate that the proposed approach has a better performance in comparison with traditional statistics approach.