Last updated on May 22, 2018 by Dotsquares
With a revolutionary research, the experts at the University of Illinois Urbana-Champlain and Intel have worked out a dataset that can improve the night-vision quality of smart phone cameras.
The experts were trying to find alternatives for improving the modern cameras especially that of smartphones, to take high-definition photos in darker settings. Usually, it is done by either ameliorating the quality of image sensors or by tweaking the exposure time. However, since neither of these methods is scalable or feasible for all the cases, they (the researchers) have tried to include the magic of Artificial Intelligence into the recipe.
In the beginning, the researchers tried to train a custom-made AI algorithm on a data set of thousands of edited photos, with different lighting effects. This synthetic data set, however, didn’t bring the desired results. In the end, they set themselves up for the laborious and time-consuming task of taking original photos with different lighting environments.
As per Chen-Chen, an Intel intern who was a part of the research, it took them months to gather the required data. To be precise he himself had spent two months in collecting low-light outdoor and a week in gathering similar indoor images. These images too were needed to be taken with precise measurements and long exposures, making it hard for the researchers to collect it, while also increasing the value of the data set which is now made publicly available.
In total, the researchers have taken two versions of over 5,000 photographs. One version being deliberately low-lit, and the other with longer exposure time, which notably took them more time in order for the sensors to collect enough light.
All these photos were then fed to the algorithm, and voila, the machine learned to amplify the exposure of low-light images by whopping 300 times, without the usual noise and discolouration effects!
The value of the research increases multi-fold when it was found that the algorithm can also improve the qualities of underexposed images taken with iPhone 6S. This could mean revolutionary changes in the basic components of the upcoming models of smart phones, with more compact solutions available for unprecedented results.
The research paper will be discussed in detail in an industry conference named CVPR in the month of June, with more hopes for the improvement of the night vision technologies through AI.
In the meanwhile, the researchers have made the data set and algorithm available for public use on GitHub, for innovators to further the research in the area.