From unlocking smartphones to authenticating identities for access to buildings or making payment, facial recognition technology has come to be used for person identification. Technological advancements have improved its precision, such that it is increasingly adopted in a variety of fields.

May 30, 2022

#Pioneering Technology#AI#Camera

Facial recognition works by automatically making one-on-one comparisons between images of faces preregistered in a database and those of faces captured in the place or situation in which facial recognition is to be conducted. The two images are determined to be of the same person when the degree of similarity between them is over a certain threshold value.

Facial recognition technology has advanced tremendously due to deep learning, a type of AI technology. “Neural networks” make it possible for computers to replicate the functionality of the human brain. Deep learning uses Deep Neural Networks (DNN), which are neural networks with many more layers. In facial recognition that is conducted via deep learning, the computer “pre-learns” from a massive database of facial information and generates a DNN model, which it then uses to extract from and compare with the facial features in captured images. This enables it to achieve a level of precision that exceeds human performance. The massive input of teaching data in the form of facial images of people of different races and ages train the computer until images of the same person give feature values that are close matches, making it possible to recognize faces with a high degree of accuracy.

Canon’s facial recognition technology harnesses the image analysis technologies that it has cultivated through development of digital and network cameras. With its large-scale databases and unique high-precision, high-speed AI deep learning architecture, it is capable of highly accurate facial recognition even for low resolution images and those shot from angles that were challenging to detect using existing technology.

To train its DNN model, Canon built a massive, diverse database of face images that consisted not just of normal photographs but also high angle images from network cameras installed in high locations, as well as images captured in poor conditions that are blurred or carry image noise. The training also accounted for differences in shooting environment such as backlight and different face angles, as well as changes in external appearance such as hairstyle changes and aging. In addition, it also trained the DNN model so that face images of the same person would give similar feature values, and those of other people with similar facial features would give feature values with larger differences, hence ensuring a high recognition accuracy rate even for images for which identification is difficult. For cases in which subjects are wearing masks or sunglasses, Canon collected a massive amount of image data in-house and input them as training data, enabling highly accurate comparison and recognition even for partially obscured faces. These training efforts, together with the optimized combination of numerous algorithms, enabled the realization of reliable and accurate facial recognition in various environments including walkthrough gates.

The FRVT is conducted throughout the year. Although its evaluation methodology is undisclosed, all participating vendors are fairly evaluated using NIST’s massive dataset consisting of face photos of 12 million people, making it an industry standard benchmarking test. Its 2021 test evaluated over 300 algorithms submitted by vendors from all over the world, including Russia and China. Canon’s excellent performance in these tests demonstrates the high accuracy of its facial recognition technology.

With sufficient consideration to the issues of privacy and potential bias, Canon will continue to polish its unique, Japan-developed AI-based facial recognition technology to provide image analysis solutions that work in synergy with its digital and network cameras.