Purity analysis of food raw materials using deep learning with anomaly detection algorithms

Against the background of the planned automation of the batch-wise visual sampling quality inspection of the food company Kündig Group, the application of the state-of-the-art anomaly detection models PatchCore and EfficientAD is investigated. In principle, such deep learning methods are only trained on normal image data (unsupervised) and can detect as well as localize deviations from the representation of normal data. For the training of corresponding models, a Dried Porcini Mushroom Slices Dataset (DPMSD) is created and published. The porcini mushrooms are spread out chaotically on a conveyor belt and recorded with an RGB (red, green blue) line scan camera. In order to assess the model performance, additional abnormal data with physical contaminants from production are added to the mushrooms and recorded, namely aluminum, plastic and glass fragments as well as dried onions and small nuts. It is shown that this industrial DPMSD has a higher complexity and difficulty compared to the typical benchmark datasets, resulting in significantly lower model performance. For example, a pixel-wise Area Under the Receiver Operating Characteristic (AUROC) of (82.3±1.2) % and an image-wise AUROC of (65.7±5.4)% is achieved for PatchCore as well as (74.8±3.7)% and (77.8±5.8)% respectively for EfficientAD. Especially the localization of the anomalies is improved by common techniques such as tiling or data augmentation. The nut anomaly cannot be detected by any model trained in this work because the similarity to the normal procini mushrooms is probably too high. Without this contaminant, the performance of an optimized EfficientAD model remains in a similar range, which is still considerably lower compared to the expectation based on the typical benchmark datasets. This illustrates the existing gap between academic research and industrial application and highlights the need for further application-oriented research.