Deep Learning Tone-Mapping and Demosaicing for Automotive Vision Systems.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Automated driving systems (ADS) require high dynamic range (HDR) imaging for safety in varied lighting.
• Current object detection algorithms are optimized for 8-bit images, not native HDR's higher bit-depth.
• Independent development of HDR imaging and object detection limits their combined effectiveness.

Purpose of the Study:

• To develop and evaluate novel convolutional neural network (CNN) architectures for converting high bit-depth HDR images to 8-bit.
• To optimize HDR to 8-bit conversion for improved object detection quality in ADS.
• To enhance the detectability of traffic-related objects in reconstructed 8-bit content while preserving realism.

Main Methods:

• Proposed two CNN architectures for intelligent HDR to 8-bit image conversion.
• First CNN: joint tone-mapping and noise suppression on full-color HDR input.
• Second CNN: joint demosaicing, tone-mapping, and noise suppression on raw HDR input.
• Comparative analysis against state-of-the-art tone-mapping and demosaicing methods using ADS object detection accuracy.

Main Results:

• The proposed CNNs demonstrate superior performance in object detection accuracy compared to standard dynamic range (SDR) content.
• The networks outperform existing state-of-the-art tone-mapping and demosaicing algorithms in object detection.
• Enhanced image quality and realism were observed in the reconstructed 8-bit content.

Conclusions:

• Novel CNNs effectively convert HDR images to 8-bit for improved ADS object detection.
• The proposed methods offer a significant advancement over current techniques for HDR processing in autonomous driving.
• This research contributes to safer automated driving through enhanced perception in challenging visual conditions.