Methods of learning discriminative features for automated visual inspection

At the present day, automation of visual inspection tasks is a typical engineering problem. Experts design the physical aspects of the system and devise classification algorithms based on a small sample of the material to be inspected. Much of this work is devoted to finding suitable features to discriminate wanted from unwanted material. In this report, we explore methods to automatically learn object descriptors from a suitably large sample. We focus on two types of descriptors: (a) global descriptors, which represent the object as a whole and (b) local descriptors, which focus on topical features. Apart from freeing the engineers to attend to other tasks, these methods allow non-experts to operate and reuse visual inspection systems, e.g. to inspect a different product than originally intended.