Squeeze-EnGAN: Memory Efficient and Unsupervised Low-Light Image Enhancement for Intelligent Vehicles

Area of Science:

• Computer Vision
• Artificial Intelligence
• Intelligent Transportation Systems

Background:

• Autonomous vehicles rely on sensors like RGB cameras, which perform poorly in low light.
• Existing low-light image enhancement (LLIE) methods are often costly or struggle with road scenes.
• Supervised LLIE methods require paired datasets, which are difficult to obtain for driving scenarios.

Purpose of the Study:

• To develop a memory-efficient, unsupervised LLIE method for intelligent vehicles.
• To address the limitations of existing LLIE models in adapting to road scenes and the lack of paired datasets.

Main Methods:

• Proposed Squeeze-EnGAN, a Generative Adversarial Network (GAN)-based LLIE method.
• Incorporated a fire module into a U-net architecture for reduced parameters and computational cost.
• Utilized an unsupervised learning approach, eliminating the need for paired low-light and normal-light datasets.

Main Results:

• Squeeze-EnGAN demonstrated significant memory efficiency and reduced Multiply-Accumulate Operations (MACs) compared to EnlightenGAN.
• Achieved real-time performance on embedded systems like Jetson Xavier.
• Enhanced images from Squeeze-EnGAN improved object detection accuracy for intelligent vehicles.

Conclusions:

• Squeeze-EnGAN offers an effective and efficient solution for low-light image enhancement in intelligent vehicles.
• The unsupervised approach overcomes the challenge of acquiring paired datasets for driving scenes.
• The model's ability to improve object detection highlights its potential for enhancing autonomous driving systems.