Research news

New paradigm for the detection of visual defects on product surfaces

Photo by Adonyig

Publish Date: 27.12.2021

Category: Outstanding research achievements, Interdisciplinary research, Our contribution to sustainable development goals

Sustainable development goals: 9 Industry, innovation and infrastructure (Indicators)

With the development of advanced data-based and teachable machine vision approaches, researchers from the Faculty of Computer Science and Information Technology of the University of Ljubljana are making breakthroughs in the field of automatic product inspection.

An ambitious goal of the fourth industrial revolution is to accomplish the automation of advanced and complex manufacturing processes, which require intelligent information processing and flexibility. Researchers from the Faculty of Computer Science and Information Technology (Assoc. Prof. Danijel Skočaj, PhD, Vitjan Zavrtanik, MSc (comp. science and inf. tech.), Jakob Božič, BSc (comp. science and inf. tech.), Domen Tabernik, PhD, Assoc. Prof. Matej Kristan, PhD) wish to introduce this trend in the field of visual inspection. Many computer vision systems are already used in supervised industrial processes for inspecting products and semi-manufactures, and for detecting potential defects or other deviations from the expected appearance or shape of inspected products. However, the classical development of such systems is time consuming and expensive, and also quite inflexible.

In the past year, the above researchers published a series of articles which introduce a new paradigm for the development of such systems. By doing so, they redirected the paradigm of manual development of specific solutions towards data-based development, which is more general, effective, adaptable and also cost-effective. This process is based on visual learning. In the learning phase, by observing images of good and bad specimens, the system builds a model which it then uses for its functioning.

Several approaches have been developed for this, based on deep learning. Typically, methods of supervised deep learning require large amounts of marked data. In their articles, the researchers proposed non-supervised methods, in which the visual model can be built by observing only the images of undamaged products, thus eliminating the need for time-consuming manual marking of defects in the learning phase.

They also developed a method which acts in the mixed-supervision mode of learning and can use information about defect markings, if available. In all cases in the phase of practical implementation, for a given input surface image, the developed neural networks are able to predict whether it contains a defect, and the defect is also marked in the image. The obtained results represent a breakthrough in the field of automatic product inspection and show large potential for the practical use of this method.

The articles were published in two excellent journals having an impact factor of over 7.5, which are thus ranked in the upper ten percent of all journals in their field, and at a prestigious conference having an index of h-5=184.


Figure: Examples of input images (above) and automatically marked defects (below). Author: Vitjan Zavrtanik


ZAVRTANIK, Vitjan, KRISTAN, Matej, SKOČAJ, Danijel. Reconstruction by inpainting for visual anomaly detection. Pattern recognition: the journal of the Pattern Recognition Society, Apr. 2021, vol. 112, 1A1, IF 7.740 (20/273), h-5 index: 99 [COBISS.SI-ID 49664003]

BOŽIČ, Jakob, TABERNIK, Domen, SKOČAJ, Danijel. Mixed supervision for surface-defect detection: from weakly to fully supervised learning. Computers in industry, Aug. 2021, vol. 129, 1A1, IF 7.635 (9/112), h-5 index: 64 [COBISS.SI-ID 63403523]

ZAVRTANIK, Vitjan, KRISTAN, Matej, SKOČAJ, Danijel. DRÆM - A discriminatively trained reconstruction embedding for surface anomaly detection, Oct. 2021, International Conference on Computer Vision ICCV 2021, h-5 index: 184

back to list