Cloud recognition – in all the face of the era, how to efficiently create face recognition services?

Machine learning is one of the ways to achieve artificial intelligence, which has been developed in the United States for decades. Machine learning also includes representation learning and deep learning.

Depth of the relationship between learning and presentation is just like a machine learning system and rules, the first is the birth of the rules system, its input and output are engineers and scientists are set in advance good rules, input how to get the content of the relevant results, the calculation of the input process is the path of the engineers manual setup.

Later, when machine learning emerged, it extracted features and produced similar results for certain types of inputs, eliminating the need to develop a large rule system.

As we move forward, there are fewer and fewer manual adjustments, and deep learning is basically an end-to-end process.

Face recognition

Face detection: the input is a photo, the output is the picture of one or more face rectangular box coordinates.

Face registration: The position of the features and contours of a face. Input is a rectangular face box, the output is the contour of the face of the key position. If you line up the dots you get the outline of a human face.

Attribute recognition: includes gender, age, facial expression, and more emotional attributes.

Feature extraction: is used to identify the face of the unique identification code.

Using the features of each face, face comparison, face verification and face recognition can be made. Face recognition here refers to the face of a photo into the system one by one comparison, which is its identity. The following face retrieval and face clustering, are based on the features of the face comparison.

In vivo detection is used for high security requirements, requiring work related to identity authentication and authentication. Other auxiliary verification means are used to ensure that the objects identified are real people rather than photos or videos.

We need to decode video formats when we watch TV or movies, which is one of the applications of the GPU. Although the computing power of each GPU core is not as good as that of CPU, it uses the number of cores to surpass the computing power of CPU.

With a CPU, the execution of programs is linear, but a GPU uses thousands of cores simultaneously to take advantage of its performance.

Develop and deploy using Docker

The emergence of Docker has greatly improved the efficiency of developers, from the development prototype to the continuous iteration of the product environment.

The API is described using Swagger, using programmer friendly JSON and YAML. After using it to describe the API can directly generate easy-to-read web documents, but also support online debugging, users are very convenient to use.

As shown above, this is an example of how we can extract points from a human face. Dot points are used for face stickers like in the live stream system. Here will be more involved in some technical details, such as mobile phone client in the use of face recognition technology is actually quite different from the server side. The mobile client is limited by its device size, and its GPU computing capacity is far less than the high-performance device used by the server, so the selection of algorithm is different.

Mobile phones typically employ non-deep learning algorithms or deep learning algorithms designed or optimized for mobile phones. Although deep learning has high accuracy, it has higher requirements on hardware performance than traditional algorithms, so a key trade-off between performance and effect of machine learning.

That’s all for today’s sharing, thank you!