Acquisition, Technology & Logistics Agency Applies GAN to Far-Infrared Images and Generates Color Images
ATLA found that MATLAB provides a rich set of deep learning frameworks and functions for developing new algorithms.
Key Outcomes
- Apply pix2pix (GAN) to far-infrared images to generate highly accurate color images
- Calibrate far-infrared and color image cameras to create training data
- Implement the methods for autonomous robots
Japan’s Acquisition, Technology & Logistics Agency (ATLA) conducts research using far-infrared imaging, which is less sensitive to changes in daylight and illumination. Far-infrared images can recognize the environment continuously and stably, but have problems such as low visibility and difficulty in responding to temperature changes.
ATLA used image-to-image translation, also called pix2pix, a method of generative adversarial networks (GAN) to generate highly accurate color images from infrared images. This approach has improved the visibility and object detection performance. ATLA used MATLAB® as one of the tools to support a workflow including image preprocessing and deep learning. Then, they applied the method to an automated guided vehicle (AGV) robot to enable it to move autonomously in the dark.
