High Dynamic Range Image Reconstruction from Saturated Images of Metallic Objects

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Reconstructing high dynamic range (HDR) images from low dynamic range (LDR) images is challenging, especially for metallic objects due to their reflective properties.
• Existing methods often struggle with accuracy and visual fidelity when dealing with complex material surfaces.

Purpose of the Study:

• To develop a deep neural network for direct reconstruction of HDR images from single saturated LDR images of metallic objects.
• To enhance the accuracy and visual quality of HDR image reconstruction for challenging materials.

Main Methods:

• A deep neural network, specifically a U-Net-like convolutional neural network (CNN) architecture, was designed for direct mapping from 8-bit LDR to HDR images.
• An HDR image database of metallic objects was created, and HDR images were clipped to generate corresponding LDR images for training and testing.
• The CNN comprised an encoder, decoder, and skip connections, utilizing MATLAB for algorithm construction with 32 layers and 85,900 learnable parameters.

Main Results:

• The proposed CNN method demonstrated superior performance in reconstructing HDR images from LDR inputs.
• Experimental evaluations confirmed significant improvements in reconstruction accuracy, histogram fitting, and subjective visual quality compared to existing methods.
• The U-Net-like architecture effectively preserved image resolution throughout the reconstruction process.

Conclusions:

• The deep neural network approach offers a robust and effective solution for HDR image reconstruction from single LDR images of metallic objects.
• The method significantly outperforms traditional techniques, providing more accurate and visually pleasing HDR results.
• This work advances the field of HDR imaging, particularly for applications involving reflective and metallic surfaces.