Low-Complexity Neural Network for SDR-to-HDR Inverse Tone Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep neural network-based inverse tone mapping methods for transforming standard dynamic range (SDR) content to high dynamic range (HDR) content are complex, making real-time implementation in products challenging.
Innovation Solution
A neural network architecture that concatenates arrays of SDR picture data samples with statistical representations, utilizing error metrics and back propagation for parameter updates, and applies iterative processing on blocks of samples to reduce complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If deep neural network based ITMO is used to transform SDR content to HDR content, then HDR picture quality is improved, but algorithm complexity increases
Solution Approach 1:
The neural network processes the image in blocks rather than treating the entire image at once. Each block is processed independently through the network, reducing the computational burden per processing unit while maintaining overall quality. This segmentation approach allows the complex DNN to be applied in a manageable, real-time manner.
Solution Approach 2:
The patent applies the neural network to only the necessary portions of the image data (luminance values in YCbCr format) rather than processing all color channels and full resolution data. This partial action approach maintains HDR quality where needed while reducing overall computational complexity.
2Device complexity
If traditional mapping functions are used for inverse tone mapping, then computational complexity is reduced, but HDR picture quality deteriorates
Solution Approach 1:
The patent introduces a trained neural network as an intermediary between the SDR input and HDR output. This DNN intermediary learns optimal tone mapping transformations from training data, providing superior quality compared to traditional mapping functions while maintaining acceptable computational complexity through efficient network architecture and block processing.
3Productivity
If deep neural network based ITMO is implemented in real-time products, then HDR content generation capability is improved, but implementation difficulty increases due to complexity
Solution Approach 1:
By segmenting the image into blocks for processing, the patent enables parallel processing capabilities that can be efficiently implemented in real-time hardware or software. This segmentation reduces the data throughput requirement for the DNN, making real-time implementation feasible in consumer products.
Solution Approach 2:
The patent transforms the input image to YCbCr color space and processes only the luminance (Y) component through the neural network. This parameter change approach reduces the amount of data processed by the DNN by approximately two-thirds, significantly easing real-time implementation requirements while preserving HDR quality in the luminance channel.
Data Source
AI summary
A method comprising obtaining a standard dynamic range (SDR) picture data; and, applying a neural network implementing an inverse tone mapping process to the SDR picture data to obtain high dynamic range (HDR) picture data, wherein the neural network comprises a concatenation of an array of samples representing the SDR picture data to an array of samples representative of at least one statistical representation of the SDR picture data.