Due to the complexity and variations in defects, surface inspection of casting relies on human vision. However, this approach becomes unstable due to human fatigue. Additionally, the lack of a standardized communication method for surface inspection criteria impedes accurate quantification. With recent advancements in machine learning and artificial intelligence, specifically with deep learning and Convolution Neural Networks (CNN), the automation of the surface inspection for castings is becoming feasible. We have developed an image analysis method based on deep learning to automate the inspection of the cast surfaces of ductile cast iron pipes. Ductile cast iron pipe is a crucial component in social infrastructure, and we must identify defective products reliably. To reliably identify defective products, it is imperative to merge quantitative grade regression with the classification of defects, ensuring no flaws are overlooked. Images of the cast surfaces underwent preprocessing and data augmentation before being input into the CNN. We compared twelve pre-trained CNN architectures based on ImageNet, eventually utilizing fine-tuned DenseNet201. For grade regression and defect classification, we proposed an ensemble learning approach, integrating a weighted average of the regressor and classifier with weighted cross-entropy as the loss function. By adjusting the weights in the proposed method, we attained grade regression while maintaining a 100% recall rate for defect classification.
