Thesis supervisor: László Czúni
Location of studies (in Hungarian): University of Pannonia, Veszprém, Egyetem str. 10. Abbreviation of location of studies: PE
Description of the research topic:
While anomaly detection in visual patterns is an old task in image processing, there are several unsolved problems we have to face to detect unusual patterns in case of complex or 3D objects. Either statistical segment models [1], [2], template matching [3], [4], or local feature detectors are used all these depend on region detection/segmentation or template registrations since each region should be localized to be able to apply the proper models. On the other hand the training of these area specific models or the detection of relevant anomalies can also be problematic since in many cases training observations for all possible types of anomalies are not available. In recent years several new approaches were developed in computer vision and image processing for image segmentation and latent information processing. The application of autoencoders [5] for segmentation, and for denoising [6] are good examples to learn typical patterns for object classes or for filtering image noise. One new idea to anomaly detection is to apply autoencoders which are known to learn and to detect the most important features of objects required for the decoder part (D) of the network (see Figure below for our proposed idea). These features are generated by the encoder (E) represented in the so called bottleneck of the network (between the encoder E and decoder D), the reconstructed signal is:
.
In our hypothesis the different unexpected patterns of anomalies will be not represented in since not trained by normal objects. Thus applying the already trained encoders on images with anomalies the decoder will fail the generation enabling the selection of regions of anomaly: . However, it is important to clarify how this approach can handle complex visual structures, occlusions or geometric distortions. In [7] augmented autoencoders are proposed to generate latent information for pose estimation. We plan to develop a similar approach now, but our purpose is not pose estimation but finding latent representation for anomaly free patterns. Another, somewhat similar, approach is the use of General Advesarial Networks (GANs). The use of GANs for training data generation is also an interesting question since it is also known that GANs can be used for latent information detection and reproduction [8]. Considering view point and illumination changes, shadows as latent image features, the visual effects of those generated by GANs can be a solution for training image generation.
Related publications:
[1] R. Cogranne and F. Retraint, “Statistical detection of defects in radiographic images using an adaptive parametric model,” Signal Processing, vol. 96, pp. 173–189, 2014.
[2] M. Basseville and I. Nikiforov, “Fault isolation for diagnosis: nuisance rejection and multiple hypotheses testing,” IFAC Proceedings Volumes, vol. 35, no. 1, pp. 143–154, 2002.
[3] X. Zhou, Y. Wang, C. Xiao, Q. Zhu, X. Lu, H. Zhang, J. Ge, and H. Zhao, “Automated visual inspection of glass bottle bottom with saliency detection and template matching,” IEEE Transactions on Instrumentation and Measurement, vol. 68, no. 11, pp. 4253–4267, 2019.
[4] D. Buniatyan, T. Macrina, D. Ih, J. Zung, and H. S. Seung, “Deep learning improves template matching by normalized cross correlation,” arXiv preprint arXiv:1705.08593, 2017.
[5] V. Badrinarayanan, A. Kendall, and R. Cipolla, “Segnet: A deep convolutional encoder-decoder architecture for image segmentation,” IEEE transactions on pattern analysis and machine intelligence, vol. 39, no. 12, pp. 2481–2495, 2017.
[6] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P.-A. Manzagol, “Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion,” Journal of machine learning research, vol. 11, no. Dec, pp. 3371–3408, 2010.
[7] M. Sundermeyer, Z.-C. Marton, M. Durner, M. Brucker, and R. Triebel, “Implicit 3d orientation learning for 6d object detection from rgb images,” in Proceedings of the European Conference on Computer Vision (ECCV), 2018, pp. 699–715.
[8] C. Donahue, Z. C. Lipton, A. Balsubramani, and J. McAuley, “Semantically decomposing the latent spaces of generative adversarial networks,” arXiv preprint arXiv:1705.07904, 2017.
Number of students who can be accepted: 1
Deadline for application: 2020-09-30
2024. IV. 17. ODT ülés Az ODT következő ülésére 2024. június 14-én, pénteken 10.00 órakor kerül sor a Semmelweis Egyetem Szenátusi termében (Bp. Üllői út 26. I. emelet).
Login
