Image generation using only a discriminator network with gradient norm penalty

Download

10497109.pdf

Date

2022-9

Author

Yeşilçimen, Cansu Cemre

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

216
views

191
downloads

This thesis explores the idea of generating images using only a discriminator network by extending a previously proposed method (Tapli, 2021) in several ways. The base method works by iteratively updating the input image, which is pure noise at the beginning while increasing the discriminator's score. We extend the training procedure of the base network by adding the following new losses: (i) total variation, (ii) N-way classification (if labels are available), and (iii) gradient norm penalty on real examples. Our experiments show that while the total variation and N-way classification do not significantly improve the performance, the gradient norm penalty results in better generative examples and faster convergence. Combining all three modifications yield the best model. Using a small convolutional network, we achieve an FID score of 25.26 on the MNIST dataset. We demonstrate additional generation results on the EMNIST and Yale Face datasets and present scores for out-of-distribution detection on FashionMNIST, EMNIST, and KMNIST datasets.

Subject Keywords

Computer vision, Gradient penalty, Convolutional neural network, Image generation

URI

https://hdl.handle.net/11511/99493

Collections

Graduate School of Natural and Applied Sciences, Thesis

Suggestions

OpenMETU
Core

Image generation by back-propagation on input using a discriminator network Taplı, Merve; Akbaş, Emre; Department of Computer Engineering (2021-9-08) In this thesis, we propose an image generation method that only involves a discriminator network; no generator or decoder networks are required. To generate an image, we iteratively apply an adversarial attack on the discriminator by updating the input image, which is noise at the beginning, to maximize the discriminator's output score. Generated images are then used as negative examples, together with the real images as positive examples, to fine-tune the discriminator. After several rounds of generation a...
Image segmentation with unified region and boundary characteristics within recursive shortest spanning tree Esen, E.; Alp, Y. K. (2007-06-13) The lack of boundary information in region based image segmentation algorithms resulted in many hybrid methods that integrate the complementary information sources of region and boundary, in order to increase the segmentation performance. In compliance with this trend, we propose a novel method to unify the region and boundary characteristics within the canonical Recursive Shortest Spanning Tree algorithm. The main idea is to incorporate the boundary information in the distance metric of RSST with minor cha...
Representing temporal knowledge in connectionist expert systems Alpaslan, Ferda Nur (1996-09-27) This paper introduces a new temporal neural networks model which can be used in connectionist expert systems. Also, a Variation of backpropagation algorithm, called the temporal feedforward backpropagation algorithm is introduced as a method for training the neural network. The algorithm was tested using training examples extracted from a medical expert system. A series of experiments were carried out using the temporal model and the temporal backpropagation algorithm. The experiments indicated that the alg...
Position estimation for timing belt drives of precision machinery using structured neural networks KILIÇ, Ergin; DOĞRUER, CAN ULAŞ; Dölen, Melik; Koku, Ahmet Buğra (2012-05-01) This paper focuses on a viable position estimation scheme for timing-belt drives using artificial neural networks. In this study, the position of a carriage (load) is calculated via a structured neural network topology accepting input from a position sensor on the actuator side of the timing belt. The paper presents a detailed discussion on the source of transmission errors. The characteristics of the error in different operation regimes are exploited to construct different network topologies. That is, a re...
Continuous dimensionality characterization of image structures Felsberg, Michael; Kalkan, Sinan; Kruger, Norbert (Elsevier BV, 2009-05-04) Intrinsic dimensionality is a concept introduced by statistics and later used in image processing to measure the dimensionality of a data set. In this paper, we introduce a continuous representation of the intrinsic dimension of an image patch in terms of its local spectrum or, equivalently, its gradient field. By making use of a cone structure and barycentric co-ordinates, we can associate three confidences to the three different ideal cases of intrinsic dimensions corresponding to homogeneous image patche...

Citation Formats

C. C. Yeşilçimen, “Image generation using only a discriminator network with gradient norm penalty,” M.S. - Master of Science, Middle East Technical University, 2022.