Anomaly detection in manufacturing line

Industrial anomaly detection becomes challenging due to the sparsity and randomness of outlier images and environmental noise. Traditional methods suffer under such circumstances, and therefore, there must be a model with accuracy and also with the capacity for speedy classification and segmentation of the anomalies. Our solution brings into action a teacher-student network for stable classification and also an autoencoder for precise localization of the anomalies. The teacher-student network discriminates between normal and abnormal images very well, and the utilization of autoencoders reconstruct images and highlight differences. Adopting the notion of EfficientAD, we strengthen this fusion with a multi-modal approach, integrating the best of both frameworks. In addition to EfficientAD we had established a loss function against overgeneralization named loss of CDO, on our plastic-nut dataset from the state of the art named Real-IAD. To overcome overgeneralization, we created synthesized anomalies mimicking real-defect scenarios. These improvements optimized the process of anomaly detection without trading off on precise segmentation and AUPRO metric. As a result, our training process consists of three components: A teacher-student network acquired from regular images for normal representation. The same technique has been applied to computer-synthesized images with imperfections for greater sensitivity. An autoencoder with the teacher network for more accurate localization of anomalies. Unlike EfficientAD, we also did not find there to be a need for a supplementary loss for student network and autoencoder output alignment. Our data has clean backgrounds with varying illumination, and we addressed this with data augmentation techniques such as contrast and brightness. Furthermore, we added algorithm for threshold optimization based on F1-score which was needed for deploying the model in real industrial environment. In conclusion, our method classifies and segments industrial abnormalities with the best of teacher-student and autoencoder frameworks. With enhanced loss functions and training techniques, our study contributes towards unsupervised industrial anomaly detection, with a generalizable and scalable approach towards quality and defects for varying factory environments.