Although vision inspection has been applied to a wide range of industrial applications, inspection accuracy remains a challenging issue due to the complexity involved in industrial inspection. THe common method adopted in industry is to use a template image as a reference template to inspect each live image on a pixel-by-pixel basis. In this thesis, a tolerance-based method is studied to replace the template image method. The said tolerance is formed by two indices computed from an image, instead of using the whole image for inspection. To ensure an accurate tolerance one, a Neural Networks method is used to take into consideration the noise and uncertainties in the parts under inspection. To reduce training time, the Taguchi method is adopted to select a minimum number of the sample images needed for training. Once a tolerance zone is obtained, a live image is inspected against it. If the indices fall inside the tolerance zone, it is deemed as good, otherwise faulty. The inspection accuracy achieved is 94.5%. Three examples are given, one for label inspection and the other two for auto part inspection.