Lightweight Residual CNN with Four Skip Connections for Grayscale Metal Casting Defect Classification

Abstract

In industrial quality control, accurate and efficient defect detection in metal casting processes remains a critical challenge, particularly in high-throughput environments. This study investigates a lightweight convolutional neural network (CNN) architecture enhanced with residual connections for classifying grayscale images of casting products into defective and non-defective categories. Using a dataset of 7,348 grayscale images sourced from Pilot Technocast, the proposed model was evaluated against standard lightweight pretrained architectures, including MobileNet, MobileNetV2, and EfficientNetV2.

The residual CNN was trained from scratch using binary cross-entropy loss and the Adam optimizer, with additional image augmentation techniques applied to improve generalization. Results show that the residual CNN achieved the highest accuracy of 99.58%, F1-score of 0.99669, and demonstrated superior precision and recall compared to all other models. Despite having only 57,665 parameters, the model outperformed more complex architectures in both performance and deployment efficiency.

These findings underscore the potential of domain-specific, residual-enhanced CNNs for real-time defect classification tasks, particularly in resource-constrained industrial settings. The model’s balance between accuracy and lightweight design makes it suitable for integration into embedded quality control systems.