Research Article
Modeling Product Quality with Deep Learning: A Comparative Exploration
@INPROCEEDINGS{10.4108/eai.28-4-2025.2358051, author={Chirumamilla Sneha and Rayavarapu Niharika and Galla Karthik and Dega Balakotaiah and Venkata Rajulu P}, title={Modeling Product Quality with Deep Learning: A Comparative Exploration}, proceedings={Proceedings of the 4th International Conference on Information Technology, Civil Innovation, Science, and Management, ICITSM 2025, 28-29 April 2025, Tiruchengode, Tamil Nadu, India, Part II}, publisher={EAI}, proceedings_a={ICITSM PART II}, year={2025}, month={10}, keywords={casting defect detection deep learning transfer learning convolutional neural network (cnn) quality control automation}, doi={10.4108/eai.28-4-2025.2358051} }- Chirumamilla Sneha
Rayavarapu Niharika
Galla Karthik
Dega Balakotaiah
Venkata Rajulu P
Year: 2025
Modeling Product Quality with Deep Learning: A Comparative Exploration
ICITSM PART II
EAI
DOI: 10.4108/eai.28-4-2025.2358051
Abstract
This work outlines a deep learning driven approach towards automatically detecting manufacturing casting defects during processes, hoping to improve industry efficiency and quality control accuracy. Submersible pump impellers classified as non-defective or defective were trained on and tested from a grayscale image dataset to benchmark different models. Preprocessing operations such as resizing, normalization, and augmentation were used to maintain consistency and stability in the input data. A tailored Convolutional Neural Network (CNN) was initially used as a baseline model, and then transfer learning was applied using three state-of-the-art architectures: EfficientNetB0, MobileNetV2, and MLP-Mixer. EfficientNetB0 obtained the highest classification accuracy of 98.75%, followed by MobileNetV2 at 98.36%. Although MLP-Mixer achieved a relatively lower accuracy of 77 08%, it provided an alternative architecture that added to the comparative study. The models were tested using standard performance measures such as accuracy, precision, recall, and F1-score. The experimental findings indicate the superiority of transfer learning in enhancing detection performance compared to traditional models. EfficientNetB0 and MobileNetV2 came out to be particularly great options for identifying casting defects with minimum misclassification.
