DxO, a software developer that has developed a reputation in the photography industry and one of the global leaders in image processing technology, is announcing the immediate release of DxO DeepPRIME, an artificial intelligence technology dedicated to editing photos in RAW format. Trained with deep learning technology using the millions of photos DxO’s laboratories have analysed for more than 15 years, DxO DeepPRIME employs cutting-edge artificial intelligence and drastically improves digital noise reduction while also delivering more effective demosaicing. The resulting photo quality is spectacular, especially for photos taken in low light conditions, with small pixels, or with early-generation cameras.
“We’ve always thought that nothing could extract pertinent information from a noisy image better than the human brain,” explained Wolf Hauser, Scientific Director at DxO. “It is fascinating to see that AI is now even better at the job!”
DxO DeepPRIME: State-of-the-art Neural Networks
DxO DeepPRIME AI belongs to a class of neural networks call convolutional neural networks, the structure of which is directly inspired by neuroscientific research on the human brain. The computer is allowed to determine the values of millions of parameters within the network, hence the term automatic learning, based on a vast database of carefully chosen input and output example images.
DxO DeepPRIME is Trained Using a One-of-a-kind Database of Calibration Images
Since its founding in 2003, DxO has earned an international reputation after measuring and calibrating hundreds of cameras and thousands of lenses using a protocol that is more extensive than any other in the industry. Its laboratories have measured the amount of distortion, vignetting, chromatic aberrations, loss of sharpness, and digital noise present in RAW images for each equipment combination and every situation with an extraordinary level of precision. Such in-depth knowledge of the performance and defects of each piece of photography equipment, which now totals to more than 60,000 different combinations, has resulted in a database of several million thoroughly calibrated images, each one containing several thousands of characteristic points. This database ultimately provided the DxO DeepPRIME network with several billion samples to learn from.
Revolutionary Denoising and Demosaicing Performance
Based on these billions of datasets, the DxO DeepPRIME artificial intelligence technology taught itself the two fundamental operations for editing any RAW image file—demosaicing and denoising.
Demosaicing consists of recreating complete colours based on the pixels from a photo sensor that initially only receive one of three basic colours. Denoising involves correcting digital noise, which produces granulation and coloured dots that significantly degrade the quality of photos taken in low-light conditions. Using a traditional approach, these two operations are performed separately, with each one introducing artifacts that degrade the quality of the other. Through deep learning, DxO DeepPRIME uses a holistic approach that combines demosaicing and denoising in the same step, resulting in exceptional performance.
“There are a number of applications for DxO DeepPRIME,” explains Jérôme Ménière, CEO of DxO. “In the field of photography, DxO DeepPRIME creates new creative opportunities by helping expert photographers overcome digital noise or a lack of light. For amateur photographers, DxO DeepPRIME compensates the drawbacks of the smaller sensors on compact and bridge cameras, which can typically only take high-quality images in bright light. Lastly, it can also be used to breathe new life into photos taken by early-generation digital cameras.”
In addition to photography for the general public, DxO is also exploring ways to apply DxO DeepPRIME in the medical, surveillance, and automotive industry.