[Paper Review] High dynamic range image forensics using cnn

Yongqing Huo, Xiaofeng Zhu|arXiv (Cornell University)|Feb 28, 2019

Digital Media Forensic Detection13 references5 citations

TL;DR

This paper proposes the first deep learning-based method for high dynamic range (HDR) image forensics using a convolutional neural network (CNN) to distinguish HDR images synthesized from multiple LDR images versus those generated via inverse tone mapping from a single LDR image. The method learns statistical特征 changes through a custom CNN architecture, achieving superior performance over traditional statistical methods in source identification tasks.

ABSTRACT

High dynamic range (HDR) imaging has recently drawn much attention in multimedia community. In this paper, we proposed a HDR image forensics method based on convolutional neural network (CNN).To our best knowledge, this is the first time to apply deep learning method on HDR image forensics. The proposed algorithm uses CNN to distinguish HDR images generated by multiple low dynamic range (LDR) images from that expanded by single LDR image using inverse tone mapping (iTM). To do this, we learn the change of statistical characteristics extracted by the proposed CNN architectures and classify two kinds of HDR images. Comparision results with some traditional statistical characteristics shows efficiency of the proposed method in HDR image source identification.

Motivation & Objective

• To address the lack of deep learning approaches in HDR image forensics.
• To identify the source of HDR images—specifically, whether they were generated from multiple LDR images or via inverse tone mapping from a single LDR image.
• To improve the accuracy of HDR image source identification using end-to-end deep learning.
• To analyze and classify changes in statistical characteristics of HDR images using a CNN-based framework.

Proposed method

• A custom convolutional neural network (CNN) architecture is designed to extract and learn statistical feature changes in HDR images.
• The network is trained to classify HDR images into two categories: those generated from multiple LDR images and those produced via inverse tone mapping (iTM) from a single LDR image.
• Statistical characteristics of HDR images are learned and compared across the two classes using the CNN's hierarchical feature extraction capability.
• The model is trained and evaluated on a dataset of HDR images with known generation sources, using supervised classification.
• The method leverages the hierarchical representation learning of CNNs to detect subtle differences in image statistics not easily captured by handcrafted features.
• Performance is evaluated against traditional statistical feature-based methods to demonstrate superiority in classification accuracy.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively distinguish HDR images generated from multiple LDR images from those created via inverse tone mapping of a single LDR image?
• RQ2 How do the statistical characteristics of HDR images differ between multi-LDR and single-LDR generation sources, and can these differences be effectively captured by a CNN?
• RQ3 Does a CNN-based approach outperform traditional statistical feature-based methods in HDR image source identification?
• RQ4 What level of performance gain does end-to-end deep learning offer over classical forensic techniques in HDR image forensics?

Key findings

• The proposed CNN-based method achieves higher classification accuracy than traditional statistical feature-based approaches in distinguishing HDR image sources.
• The model successfully learns and exploits subtle statistical differences in HDR images generated through different processes.
• The use of deep learning enables automatic and robust feature learning, reducing reliance on manually engineered statistical features.
• The results demonstrate the feasibility and effectiveness of applying CNNs to HDR image forensics, marking a novel application of deep learning in this domain.