Lightweight unsupervised image enhancement based on bilateral grid

Visual perception serves as the primary sensory modality for autonomous robots navigating and interacting with complex environments. However, robotic vision systems frequently encounter challenging illumination conditions—such as low-light indoor corridors, high-contrast outdoor scenes, and backlit industrial workspaces—that severely degrade the quality of captured images, directly compromising downstream robotic tasks including object detection, semantic segmentation, and visual SLAM. Deep learning-based methods have achieved remarkable success in image enhancement, but traditional approaches often necessitate paired datasets for training, incurring high costs for acquiring paired pre- and post-optimized images in real-world robotic scenarios. This paper introduces a lightweight image enhancement network named Cycle Bilateral Generative Adversarial Networks (CBGAN), specifically designed for real-time visual perception in robotic systems. Based on an unsupervised cyclic generative adversarial network architecture, CBGAN optimizes images by learning to predict both local and global affine transformations within the bilateral space in a low-resolution stream, and subsequently applying these transformations to the full-resolution domain. Specifically, CBGAN employs adversarial training, incorporating inverse mapping and cycle consistency loss, to ensure high-quality enhancement within an unsupervised training paradigm. In unsupervised scenarios, CBGAN outperforms existing state-of-the-art methods. Our network processes a 4 K image in tens of milliseconds. Compared with the conventional CycleGAN employing ResNet generators, our method achieves an approximate 95.8% reduction in parameter count and operates over 170 times faster, demonstrating immense potential for real-time deployment on resource-constrained robotic platforms.