[Facial Recognition] Development & Background

 by Han-Ru Xie

Early projects- machine learning and algorithms

Facial recognition refers to the technology capturing static images or dynamic sequences in a certain scene and compare to them with data on human faces in a database to identify the identity of one or more people in the target scene.

The technology in the early days were used based on traditional machine learning to capture geometric features of human faces, which include the distance between eyes and the contour of head and nose. The algorithm analyzes the relative position of these features, such as the distance between the corners of eyes and the tip of nose. Yet, it was quite difficult to accurately locate the detailed features, which are too rough and simple to contribute to high accuracy.

Based on the features collected, the next step was to build two-dimensional images. The features collected at this step are no longer semantic information, but underlying physical features, such as image grayscale and transform coefficients. However, face features come with several layers, making it hard for identification and training of machine learning models.

Traditional machine learning relies on analyzing facial features

Breakthrough- deep learning algorithm

Around 2013, as deep learning algorithms made huge progress than that of traditional machine learning in handwritten text recognition, coupled with the prevalence of parallel computing using GPUs, researchers have started to develop facial recognition algorithms using deep learning.

With Labled Faces in the Wild (LFW), deep learning algorithms solve the challenges faced using traditional machine learning when it comes to analyzing two-dimensional images. The results, for the first time, were 97% accurate. Since deep learning algorithms were adopted, the facial recognition systems developed by different companies have become more standardized and consistent.

The most common application of deep learning in facial recognition is to extract distinguishing features from the input face images from the layered convolutional neural network (CNN) model. Based on the data, the algorithm calculates the cosine distance for identify recognition. To achieve this goal and improve performance of a deep learning model, a large amount of data is the key.

(圖片來源:)
After 2013, deep learning has been widely used in facial recognition

Conclusion

Knowing the history of facial recognition, people can see this as an opportunity to think about how to improve the technology, while at the same time, better manage and make use of AI in the future, so that the benefits can be maximized for all humans.

【vMaker台灣自造者粉專】
>> https://lihi1.cc/kLUKF
⭐新奇有趣的科技新知、生活手作內容
⭐提供創客原創文章及活動資訊

 

分享到社群

vMaker編輯

The share account of vmaker editors. Send your work to us:contact@vmaker.tw