Abstract
Intelligent defect detection is gaining popularity in companies, in lieu of manual quality check. An automated visual inspection system can be used to check high-quality items while minimizing the strain on the labor. The shell casting process of the pump impeller is prone to surface defects like extra projection, shrinkage, etc., due to high operating temperatures. CNN is a deep learning technique employed for performing quality check which prefers image as an input and classifies it according to the features. An organization's original dataset was used which consisted of 7348 images. The dataset was divided into four categories, i.e., extra projection, shrinkage, multiple defects, and ideal workpiece. The 90% of original dataset (i.e., 6633 from 7348) was split into 70:30 for training and validation, respectively. Augmentation was done to increase the dataset size to 12,000 images in order to avoid overfitting. To assess the capability of architecture for an unobserved data (i.e., testing), 300 images were randomly chosen from the remaining 10% of the original dataset (i.e., 715). MobileNet V2, ResNet-18 and general model were chosen as the CNN architecture. Out of the three models, MobileNet V2 portrayed promising results with the highest test accuracy of 98.17% followed by the ResNet-18 and the general models which attained the test accuracies of 97.58% and 93.58%, respectively. A graphical user interface (GUI) was developed to assist workers in addition to the suggestion of remedial measures for the defects.