Séminaire (organisé par l’équipe CID)


Doctorant au laboratoire Roberval, UTC

Deep Generative Models and AutoEncoders to robustly detect abnormal situation – Application to manufacturing defects detection

Mardi 30 janvier 2018 à 10h15 en GI042

Résumé : The manufacturing industry is always trying to automate many of its engineering tasks and workflows. Applying Machine/Deep Learning in such an industrial context is challenging for different reasons : Contrary to the GAFAM (Google, Apple, Facebook, Amazon, Microsoft) and BATX (Baidu, Alibaba, Tencent and Xiaomi), the manufacturing industry use on a day-to-day basis highly complex data which comes, most of the time, in proprietary formats (e.g. 3D Models). Moreover, data are not available in large quantities (impossibility to rely on a big data scale) and most of the time in insufficient quantities to train any Deep Learning model to perform any computer vision task. However, Machine Learning can still be applied using various tips and tricks, frequently with good results. Transfer Learning is one effective solution, nevertheless this presentation will not focus on the latter which is already well covered in the literature [Y. Bengio, 2012] – [M. Oquab, 2014]. Inspired by research in anomaly detection for neuro-imagery [K. Wang, 2016] - [T. Schlegel, 2017] and for patrol robots [W. Lawson, 2017], we have studied the effectiveness of Deep Generative Networks and Deep Auto Encoder, namely GANs (Generative Adversarial Networks) and VAE (Variational AutoEncoders). This presentation aims to present the specificities and basics of GANs/VAEs and an overview of the whole process we developed with a limited amount of data (1000 images) and obtain 92% of accuracy and 85% of IOU (Intersection over Union) for the machined surface defect localisation. The proposed approach is able to adapt itself without any change in the structure or hyper-parameter settings to different industrial situations offered by the challenge dataset provided by DAGM 2007. In summary, the strength of the proposed solution is the following : Highly adaptable with limited re-engineering costs, reduced training time (+/- 3 hours on one GPU), Weakly Supervised Approach thus having a very reduced deployment cost.


FR SHIC 3272

Collegium UTC/CNRS