Enhancing Neural Autoregressive Distribution Estimators for Image Reconstruction
This addresses image reconstruction from limited observations for computer vision applications, but is incremental as it adapts existing autoregressive methods.
The paper tackles image reconstruction from sparse pixel observations by proposing a generalized ConvNADE model for real-valued/color images and comparing random vs. low-discrepancy pixel patches. Experiments show that uniform pixel coverage improves reconstruction fidelity and test performance on benchmark datasets.
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.