Methods for automatic anomaly detection in tomographic images of food and wood products

Anomaly detection refers to the problem of finding patterns in data that deviate from the expected behaviour. This technique is applied in the food industry to find foreign bodies in the products and discard them from the production line. The detection can be achieved via algorithms that use tomographic images. These 3-dimensional images are obtained using x-rays, allowing a complete view of the different densities inside a product. The detection algorithms, depending on the characteristics of the considered product, require an adaptation step and a fine-tuning of hundreds of parameters. In this work, carried out at Microtec, the anomaly detection problem is addressed via deep learning techniques: a custom autoencoder model is proposed to avoid the high amount of time required to adapt the traditional algorithms to newproducts. This autoencoder implements a memory of the latent space that helps the model to remove the anomalies in the output images. Afterward, the applicability of the derived model is also evaluated for wood tomography images, evaluating its versatility in different fields.