Evaluation of Computer Vision Techniques for Quality Inspection in Casting Manufacturing Process

Modern casting manufacturing demands have outgrown the capabilities of many traditional quality-inspection processes. Integrating technologies, such as Artificial Intelligence (AI) into the Internet of Things (IoT) has shown a way to automate and improve defect detection in manufacturing industries. However, with several AI models available for these tasks, choosing the most appropriate model for specific quality inspection applications remains an open challenge. This study evaluates and compares the effectiveness of two computer vision models, a custom Convolutional Neural Network (CNN) and a transfer learning model (MobileNetV2), for detecting defects in casted submersible pumps’ impellers. The models were trained on a dataset containing images of defective and non-defective casted submersible pump impellers. Prior to training, data augmentation techniques were applied to enhance the models’ performance by increasing the size and diversity of the training dataset. The performance of both models was evaluated using precision, recall, accuracy, F1-score, and inference time as evaluation metrics. The results of the evaluation showed that the precision, recall, accuracy, F1-score, and average inference time were 99.69%, 98.04%, 98.70%, 98.85%, and 0.2532 seconds per image, respectively, for the Custom-CNN model, while for MobileNetV2 they were 98.69%, 99.34%, 98.86%, 99.01%, and 0.3107 seconds per image, respectively. Statistical analysis showed no significant performance difference (5% significance) between the models’ classification performance. Both models are effective for defect detection in cast pump impellers. However, the custom CNN had a notably faster inference, implying that it is more suitable when speed is critical.