Irregular shapes classification by back-propagation neural networks

This paper proposes a back-propagation neural network approach to classify irregular shapes by their convex hulls and brightness in automated production lines. An image-based defect detection and classification system is established, which would examine the quality of rolls of aluminum foil and determine the type of defects, such as bolt, fracture, scratch, or spot on the aluminum foil. The developed approach is capable of performing image acquisition, image processing, defect detection and recognition, and, subsequently, the classification of the aluminum foil sheets by a back-propagation neural network. In order to verify the effectiveness of the developed approach, ten-fold cross-validation is used. The experimental results show that, using a small number of training iterations, the average accuracy rate of classification reaches 96.4%. Thus, the developed approach can be used to replace manual visual inspection for process control and process improvement, which significantly reduces the cost of labor and increases the consistency of product quality.