研究目的
To propose a deep learning-based method for inverse tone mapping to generate high-quality HDR images from single SDR images, optimized for HDR displays.
研究成果
The proposed CNN-based inverse tone mapping method effectively generates high-quality HDR images from single SDR images, outperforming conventional methods in objective metrics. It is applicable to high-resolution images and can be adapted for HDR video. Future work should address limitations in high luminance areas.
研究不足
The method does not perform well in very high luminance areas such as sunlight due to insufficient data in the LUCORE dataset. There may be banding artifacts in smooth gradation areas if too many training iterations are used.
1：Experimental Design and Method Selection:
A CNN-based method is designed with four fully convolutional layers using ReLU activation, except the last layer, and zero padding to maintain input-output size. The loss function is Mean Squared Error (MSE) minimized via stochastic gradient descent (SGD).
2：Sample Selection and Data Sources:
The LUCORE dataset is used, consisting of HDR and SDR images in TIFF 16-bit format, conforming to BT.2020 and ST 2084 standards, with 254,200 SDR image patches for training.
3：List of Experimental Equipment and Materials:
Not specified in the paper.
4：Experimental Procedures and Operational Workflow:
Input images are split into nine sub-images using mirror padding to avoid boundary artifacts, processed by the CNN, and merged. Testing is done on high-resolution (4K) images.
5：Data Analysis Methods:
Performance is evaluated using Peak Signal-to-Noise Ratio (PSNR) and HDR-VDP-2 metrics for objective quality assessment.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
