Sabri, KhierGaceb, Mohamed2025-12-10202502683768https://link.springer.com/article/10.1007/s00170-025-16390-1https://dspace.univ-boumerdes.dz/handle/123456789/15861This study introduces a vision-driven defect detection system leveraging advanced deep learning architectures to enhance quality control in metal manufacturing. The system integrates four models: the VGG-19 Network, an Attention-Augmented Convolutional Neural Network (CNN), a Vision Transformer (ViT), and a Convolutional Autoencoder. These models were trained and evaluated on a dataset of annotated high-resolution images of impellers acquired under controlled industrial conditions. Among the classifiers, the VGG-19 Network achieved an overall classification accuracy of 100% on this dataset, with F1-scores of 100% for defective samples and 99% for acceptable ones. Grad-CAM visualizations were used to highlight the critical regions influencing the VGG-19 network’s classification decisions. The Attention-Augmented CNN achieved an accuracy of 98.17%. The ViT model yielded F1-scores of 99.42% on the training set and 100% on the validation set, with a 0% false positive rate and a 1.54% false negative rate. The Convolutional Autoencoder enabled unsupervised segmentation by detecting pixel-wise reconstruction errors that indicate anomalies. Each of the four models exhibits an acceptable level of performance with respect to both computational efficiency and deployment feasibility for real-time anomaly detection in manufacture processesenComputer VisionMachine LearningObject RecognitionVirtual and Augmented RealityVisual systemArticle