Deep Learning Binary-Classification Model for Casting Products Inspection

Abstract

It is imperative for a manufacturing process to not only have a quality assurance system, but that system should also be a very efficient one. While conventional methods have always involved the human element in quality control, their inefficiency and economic liability have always been a cause of concern. An Image Classification inspection system has the capability of minimizing cost factors and can also provide a near-perfect efficient quality check. This paper focuses on developing Convolutional Neural Network (CNN) architecture to scrutinize defects in casting products. The CNN is trained with a dataset of grey-scaled images of top-view of a casted submersible pump impeller, and the trained model gives a very encouraging result in detecting various surface manufacturing defects and ultimately classifies the input image of the casted products manufactured as acceptable or unacceptable for a quality check process. A comparative study has also been done with a pretrained Xception model to analyze the performance of results achieved by our proposed model