Yesterday afternoon, Huang Jiangji, co-founder of Xiaomi Technology, announced on Weibo that Xiaomi’s new face detection algorithm has achieved accuracy in the face detection accuracy rate of FDDB (Faster RCNN based on Bootstrapped by Hard Negative Mining). One's results. Lei Jun forwarded this Weibo for the first time.
According to the FDDB official website, this new detection algorithm based on deep convolution detection network was developed by Dr. Wan Haohua and his team. The algorithm is based on a deep convolution detection network and recognizes the position and size of the human face by learning the features of human faces and non-human faces.
The following figure is the FDDB official face detection accuracy recall rate curve diagram, the horizontal axis represents the number of false detections face, the vertical axis is the detection rate. The steeper the curve, the better the detector performance. The results from the data image show that Xiaomei's corresponding line recognition detection rate is significantly higher than other research teams.
Face detection is not a new technology, and it has been widely used in mobile phones, including the iPhone. Through face detection, the camera of the mobile phone can accurately capture the position of the face and identify the specific target. On June 5, 2015, MIUI launched a new feature called Face Album, which can use computer vision technology to automatically sort photos of cloud albums by faces. According to Xiaomi, Xiaomi's advantage in face detection is that he has a large amount of user photo data, which makes it possible to have a sufficient amount of map data training.
According to official statistics of millet, currently users use Xiaomi cloud album to store 120 million photos each day, with a total volume of over 50 billion. Under the premise of ensuring sufficient data, Xiaomi will continue to optimize the algorithm. Xiaomi said that they will gradually replace the old algorithm after they continue to optimize and test the new algorithm for face detection.
What is the processing of face detection and recognition? Where are the core difficulties? What are the new application scenarios? Lei Feng network (search "Lei Feng network" public number attention) later feature articles will give detailed reports. stay tuned!