Comparative Evaluation of Pretrained Convolutional Neural Networks with Attention Enhancements for Defect Detection in Injection-Molded Products

Recent advances in machine and deep learning have paved the way for automated quality control in injection molding. However, existing approaches still face challenges such as insufficient performance evaluations, scarce training data, and limited generalization. In this study, a comparative evaluation of commonly used pretrained Convolutional Neural Networks, including InceptionV3, ResNet-101, ResNet-50, and VGG16, for defect detection was carried out. It was shown that, InceptionV3 outperformed the other models in test accuracy and convergence behavior, under consistent hyperparameter settings and fine-tuning strategies. Building on this Squeeze-and-Excitation (SE) and Convolutional Block Attention Module (CBAM) were integrated into InceptionV3, forming InceptionV3-CBAM-SE, to enhance feature representation. The proposed framework was tested on three diverse injection molding datasets to demonstrate generalized application and industrial adaptability. Compared to the baseline InceptionV3, the InceptionV3-CBAM-SE achieved accuracy gains in excess of 66%, 61.67%, and 47.62% on the three diverse injection molding datasets, reaching final accuracies of 83%, 97%, and 93%. These results demonstrate the effectiveness of the proposed model in improving defect detection performance in injection molded products.