“The DeepExposure AI method is said to use algorithms to further enhance smartphone’s photography”
Xiaomi is all set to embark a new image processing method called DeepExposure AI on its future smartphones. The Chinese OEM has recently released a paper that aims to solve common issues of exposure, light, and details in smartphone’s photography. The company’s new method is said to use algorithms to segment an image into multiple “sub-images,” instead of one, and adjust their exposure separately for better results.
The illustration of DeepExposure AI was shown in an image (below). As you can see, the buildings are white, the clouds are white, and the overexposed sky is also white. The algorithms do a great job at restoring details in the final image.
“The accurate exposure is the key of capturing high-quality photos in computational photography, especially for mobile phones that are limited by sizes of camera modules,” reads Xiaomi’s DeepExposure AI research paper. It further added that “Inspired by luminosity masks usually applied by professional photographers we’ll develop a novel algorithm for learning local exposures with deep reinforcement adversarial learning.”
As per the research paper, the DeepExposure AI divides the whole image into image segmentation and ‘action-generation’ stage. The latter concatenate low-resolution, sub-images, and direct fusion of the shot you clicked, and then process it by a policy network to compute different exposures – local and global – to evaluate the overall picture quality. The neural networks are based on Generative adversarial network (GAN) that, in a nutshell, can produce samples and distinguish between a generated samples and real-world samples.
The DeepExposure AI is built on Google’s TensorFlow framework on Nvidia P40 Tesla GPU. Xiaomi claims that its research team has used the images from MIT-Adobe FiveK Dataset to teach the AI the definition of a perfect shot. The catalogue consists of unedited RAW images as well as some photos retouched by experts. The network is said to work on low-resolution images and is tasked is to come up with the best parameters for classic image filters.