This paper describes the development of a visual inspection system for detection of blemishes in manufacturing parts. Initially, median filtering is used to improve the quality of the picture, next the Sobel operator is applied to locate the edges in the image, followed by Otsu's thresholding procedure that yields the binary edge image. The image is then decomposed into a number of non-overlapping, coarse resolution images, using the technique of coarse coding, and fed into a second-order neural network equipped with a built-in feature extraction mechanism. Finally, the performance of the system is evaluated and a study of its robustness to various types of noise is performed.