Defect.AI is a research project funded by the FFG.

Motivation And Key Objectives

The automatic and reliable detection of defects occuring in production is relevant for many industry sectors and is a key pillar of Smart Manufacturing and Industry 4.0. Usually done with visual methods, it has numerous benefits: Higher customer satisfaction due to increased product quality, minimisation of costly warranty claims, higher reputation of the product and producer and many more. In the same way, the automatic detection of defects in multimedia content (e.g. artifacts in digital video, dust spots in film material) is highly desired, as it allows for full automation of quality assessment in production and distribution (in times of increasing diversity of target channels to be served), as well as for cost-effective mass digitisation and automatic restoration of large visual media collections (e.g, broadcast archives).

Unfortunately, the adoption of deep-learning based AI methods in both the manufacturing and media sector is still lacking. The main obstacle is clearly the lack of annotated data for the effective training of a model tailored for the detection of a certain defect (e.g. surface defects on manufactured goods, dust or other defects on film material, visual playback defects in digitised video or film media). Generally, at least several thousands or ten-thousands of samples annotated by a domain expert are needed to train a robust defect detection model. Having a dataset of insufficient size for training usually leads to a model which performs poorly in practice. But in many real-world applications in manufacturing and media, it is not possible or not viable to gather and annotate such large amounts of training data. This may be due to the prohibitive cost of human expert annotation, ownership/copyright issues on the data, or simply not having enough content showing defects of a certain kind. On the other hand, often sophisticated hand-crafted detectors are currently employed, which have been specifically tailored to the characteristic appearance of a certain defect (spatial and temporal) by an expert and have been developed before the deep learning era.

Motivated by these shortcomings for the adoption of deep learning methods in manufacturing and media, Defect.AI aims at drastically reducing the data needed for the effective training of a defect detection model, which will lead to a much broader and faster take-up of these methods. Specifically, we aim for effective training of a defect detection model with only a fraction of the usual dataset set size. Our goal is to train a model using only 10% of the usual size, which has a similar performance to the model trained on the full dataset. This means that the effort for the human annotation of the dataset and the training itself are reduced ten times, leading to significant savings in both costs and training time. For many defects, this significant reduction of the dataset size needed for training even makes the usage of deep learning methods possible for the first time. Furthermore, Defect.AI aims at capturing expert knowledge embedded in existing sophisticated hand-crafted detectors, which will help with the transition of a quality inspection process from hand-crafted methods to deep learning based models.