Introspective Energy Models: Outlier detection by failure of inpainting

Kinder, Lukas (2023) Introspective Energy Models: Outlier detection by failure of inpainting. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text mAI_2023_KinderL.pdf Download (2MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (129kB)

Abstract

Detecting outliers in images is a challenging task for deep-learning models, if the training data contains little or no examples of outliers. Generative adversarial neural networks (GANs) or autoencoders can be used for this purpose, but usually require a lot of training data. In the MVTec dataset, pure deep-learning models are outperformed by distance based models, which identify outliers based on their high distance from samples in the training set. However, a problem with them is that they rely on pretrained convolutional kernels and are not very explainable. This thesis presents the idea of an introspective energy model, that measures the success of inpainting to detect local anomalies in images. The approach involves a two-stage process. First, convolution is applied to generate feature maps of the image. Subsequently, the features of image regions are predicted using the features of surrounding regions. Inaccurate predictions indicate an outlier. This mechanism initially used Bayesian networks but was later refined using neural networks. The results demonstrate that introspective energy models can outperform the state of the art for certain object categories of the MVTec dataset. In a second experiment convolutional kernels pretrained on the ImageNet dataset were used in an attempt to improve the model further. However, in this case the model performed worse than the state of the art. This work is relevant because it uses a new mechanism to train convolutional layers

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Verheij, H.B. and Grzegorczyk, M.A.
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	02 Aug 2023 09:16
Last Modified:	02 Aug 2023 09:16
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/31039

Actions (login required)

View Item