Massively parallel VLSI-implementation of a dedicated neural network for anomaly detection in automated visual quality control

In this work we will present the VLSI-implementation of a dedicated neural network architecture which we have developed in prior work for anomaly detection in automated visual industrial quality control. The network, denoted as NOVAS performs a filtering of inspection images and highlights defects or anomalies in an isomorphic image representation, allowing the detection and localisation of faults on objects. Training of NOVAS is achieved by simply presenting a set of tolerable objects to the network in a single sweep. NOVAS works with single and with multichannel image representations. The processing principle of NOVAS is closely related to nearest neighbor and hypersphere classifier approaches. We have designed an ASIC for the efficient implementation of the nearest neighbor search. Based on that ASIC we will present an architecture of a modular massively parallel computer suited to meet the real-time constraints of manufacturing processes. Further we will report on the status of a prototype system which is close to completion.