Enhancing Neural Autoregressive Distribution Estimators for Image Reconstruction

TL;DR

This paper improves neural autoregressive models for image reconstruction by exploring pixel sampling strategies, including quasi-Monte Carlo inspired methods, to enhance the quality of image completion tasks.

Contribution

It introduces a generalized convolutional autoregressive model for real-valued images and analyzes the impact of different pixel sampling methods on reconstruction quality.

Findings

Uniform pixel sampling improves reconstruction fidelity.

Quasi-Monte Carlo inspired pixel patches enhance test performance.

Design choices in pixel selection significantly affect model accuracy.

Abstract

Autoregressive models are often employed to learn distributions of image data by decomposing the $D$-dimensional density function into a product of one-dimensional conditional distributions. Each conditional depends on preceding variables (pixels, in the case of image data), making the order in which variables are processed fundamental to the model performance. In this paper, we study the problem of observing a small subset of image pixels (referred to as a pixel patch) to predict the unobserved parts of the image. As our prediction mechanism, we propose a generalized version of the convolutional neural autoregressive distribution estimation (ConvNADE) model adapted for real-valued and color images. Moreover, we investigate the quality of image reconstruction when observing both random pixel patches and low-discrepancy pixel patches inspired by quasi-Monte Carlo theory. Experiments on benchmark datasets demonstrate that, where design permits, pixels sampled or stored to preserve uniform coverage improves reconstruction fidelity and test performance.