Representing Camera Response Function by a Single Latent Variable and Fully Connected Neural Network

TL;DR

This paper introduces a novel camera response model using a single latent variable and a fully connected neural network, achieving high accuracy and faster calibration compared to existing models.

Contribution

It proposes a new high-performance, efficient camera response model with a single latent variable and neural network architecture search, trained via unsupervised learning.

Findings

Achieved state-of-the-art CRF representation accuracy.

Nearly twice as fast as current models during calibration.

Effective latent distribution learning constrains the model.

Abstract

Modelling the mapping from scene irradiance to image intensity is essential for many computer vision tasks. Such mapping is known as the camera response. Most digital cameras use a nonlinear function to map irradiance, as measured by the sensor to an image intensity used to record the photograph. Modelling of the response is necessary for the nonlinear calibration. In this paper, a new high-performance camera response model that uses a single latent variable and fully connected neural network is proposed. The model is produced using unsupervised learning with an autoencoder on real-world (example) camera responses. Neural architecture searching is then used to find the optimal neural network architecture. A latent distribution learning approach was introduced to constrain the latent distribution. The proposed model achieved state-of-the-art CRF representation accuracy in a number of benchmark tests, but is almost twice as fast as the best current models when performing the maximum likelihood estimation during camera response calibration due to the simple yet efficient model representation.