Single-Image Super-Resolution Using a Sparsely Gated Mixture of Experts

Author(s)

Zachary Huang

School Name

Spring Valley High School

Grade Level

10th Grade

Presentation Topic

Computer Science

Presentation Type

Non-Mentored

Abstract

Single-image super-resolution is the process of generating a high-resolution output image from a single low-resolution input. Deep convolutional neural networks have been successfully applied to this task. The purpose of this experiment was to determine if applying the sparsely-gated mixture-of-experts architecture can enhance the performance of convolutional neural networks for super-resolution. It was hypothesized that, due to model variation, the mixture-of-experts model would achieve a higher quality of super-resolution than a single network and would not be as computationally expensive. A mixture-of-experts model for super-resolution was developed using Tensorflow, and each expert was a single convolutional neural network. The total number of experts varied from 1 to 8, and based on a gating network, a single selected expert was activated to calculate an output for each image. The model was trained and tested for 2x image super-resolution on low-resolution and high-resolution image pairs of size 16x16 and 32x32, respectively. Its performance was measured by mean squared error (MSE) and the structural similarity index (SSIM). It was found that, compared to the base model (one expert), increasing the number of experts introduced a decrease in both the quantitative and qualitative performance of the model. The models with two, four, six, and eight experts all had statistically significant decreases in SSIM, with p = .019, p = .009, p <.0001, and p <.0001, respectively. The listed models also all had statistically significant increases in MSE, each with p < .0001. The mixture-of-experts model implemented in this paper was not favorable for single-image super-resolution.

Location

Furman Hall 109

Start Date

3-28-2020 8:30 AM

Presentation Format

Oral and Written

Group Project

No

COinS
 
Mar 28th, 8:30 AM

Single-Image Super-Resolution Using a Sparsely Gated Mixture of Experts

Furman Hall 109

Single-image super-resolution is the process of generating a high-resolution output image from a single low-resolution input. Deep convolutional neural networks have been successfully applied to this task. The purpose of this experiment was to determine if applying the sparsely-gated mixture-of-experts architecture can enhance the performance of convolutional neural networks for super-resolution. It was hypothesized that, due to model variation, the mixture-of-experts model would achieve a higher quality of super-resolution than a single network and would not be as computationally expensive. A mixture-of-experts model for super-resolution was developed using Tensorflow, and each expert was a single convolutional neural network. The total number of experts varied from 1 to 8, and based on a gating network, a single selected expert was activated to calculate an output for each image. The model was trained and tested for 2x image super-resolution on low-resolution and high-resolution image pairs of size 16x16 and 32x32, respectively. Its performance was measured by mean squared error (MSE) and the structural similarity index (SSIM). It was found that, compared to the base model (one expert), increasing the number of experts introduced a decrease in both the quantitative and qualitative performance of the model. The models with two, four, six, and eight experts all had statistically significant decreases in SSIM, with p = .019, p = .009, p <.0001, and p <.0001, respectively. The listed models also all had statistically significant increases in MSE, each with p < .0001. The mixture-of-experts model implemented in this paper was not favorable for single-image super-resolution.