Högskolan i Skövde

This study investigates the suitability of various deep learning-based anomaly detection algorithms for automated quality inspection in industrial settings, with a focus on images of cylinder heads. Manual quality inspection has several challenges that can be addressed by automated quality inspection. Using the open-source library Anomalib, nine anomaly detection algorithms are evaluated based on their discriminative ability, peak video random access memory usage, and time efficiency. The experiments are conducted using real-world manufacturing data provided by Aurobay, and models are assessed through a quasi-experimental approach using multiple performance metrics. The results showed that FastFlow achieved the highest overall discriminative ability, while DFKDE and DFM had the best time efficiency. For video random access memory, PatchCore had the highest peak usage. The findings can assist organizations in achieving reliable and cost-effective automated quality inspection systems. Directions for future work include utilizing more datasets, optimizing hyperparameters, and implementing a model into a real-world workflow.