I have a talent market

Chapter 86 Facial Recognition Technology

Technician: "The pricing of this kind of electronic product is very vague. It is very expensive to buy one, and it can be much cheaper to buy in bulk, and the internal price of the industry is different from the price of a large customer, so you can find someone in the industry to get it from a wholesaler. goods."

Zhang Xu: "It's easy, I'll give you a call later, you negotiate the price with him, first purchase 1000, and then place a reservation contract for 2000."

"Good boss!"

…………

"Boss, the research institute is running out of money!" Seeing that Zhang Xu was about to leave, Wang Hao hurriedly followed him and cried poorly.

"Why are you out of money again, didn't you just call you for the last money?"

Wang Hao: "That happened last month. I sent you an email yesterday. Didn't you see it?"

Zhang Xu had actually seen it a long time ago. This was just his subconscious question. He was bald by Wang Hao. When he heard him ask for money, he subconsciously wanted to refuse.

Seeing that Zhang Xu didn't say anything, Wang Hao thought that Zhang Xu was dissatisfied with the [Artificial Intelligence] research team. After all, for a month, just asking for money but not getting results would be unbearable for anyone, so he took him to see this one. month's labor.

"What is this?" Zhang Xu asked looking at a screen with a camera.

"Face recognition punch card machine!"

"What, punch card machine!"

"It's not a punch card machine, it's a face recognition punch card machine."

"What machine?"

"Face recognition punch card machine!"

Zhang Xu was a little annoyed, a punch card machine cost him more than 1000 million, and his trust in Wang Hao was rewarded in this way, Zhang Xu's face is very ugly now.

The technician on the side saw that Zhang Xu's complexion was not good, and Director Wang was still there clumsily introducing them, so he quickly followed their footsteps, approached Zhang Xu's other half and explained to him: "This is a computer vision research and development project. later product."

"Computer vision?" Zhang Xu turned his head when he heard a new term.

"Yes, computer vision is the ability of a computer to recognize objects, scenes, and activities in images. Computer vision techniques use sequences of image processing operations and other techniques to break down image analysis tasks into manageable pieces. block task."

Zhang Xu's expression softened when he heard this. He glanced at Wang Hao angrily, and complained to him, "Director Wang's sales ability needs to be improved. If you are asked to sell products to others, you probably won't be able to sell any of them."

Wang Hao scratched his head in embarrassment, and looked at the young man next to him and introduced to Zhang Xu: "This is Luo Cheng, he is my deputy, and his mouth is the most powerful among us, let him introduce it to you, I'd rather Concentrate on researching technology."

Zhang Xu: "Haha, it's good that you know that a high-tech company has invested 1000 million yuan to develop a punch card machine. It will make people laugh out loud."

Luo Cheng: "Director Wang is focusing on the face recognition system, not the punch card machine."

Wang Hao: "Yes, yes, I have been emphasizing this, and the boss thought I was talking about the punch card machine."

Zhang Xu blushed immediately, he changed the subject and said, "Let me see your high-tech time card machine."

Luo Cheng winked, and he took a step forward to demonstrate to Zhang Xu.

I saw Luo Cheng standing in front of the camera, and the monitor that was originally black suddenly lit up, jumping Zhang Xu who was observing.

Luo Cheng slightly adjusted the distance and angle, so that his face was completely displayed on the screen, and in the next second, the words "check in successfully" were displayed on the screen, and his check in information was simultaneously refreshed on the computer next to him.

Zhang Xu: "It's nothing, it's available both in foreign countries and in China! It's nothing special!"

Luo Cheng quickly explained: "The face recognition system of the punch card machine has been around for a long time, but ours is more advanced?"

"How to say?"

"Early face recognition algorithms include geometric features, template matching, local features, subspace algorithms, principal component analysis, linear discriminant analysis, and sparse representation."

"what about now?"

Luo Cheng: "Now there are basically [artificial features + classifiers] in the world. Artificial features are divided into: histogram of directional gradients, scale-invariant feature changes, and local binary patterns. The idea of ​​​​classification is to give a positive example and The sample collection of negative examples, the purpose is to find a hyperplane to separate the sample positive and negative examples, and make the point closest to the hyperplane among all points have a larger distance from the hyperplane.”

"What about you now?"

Luo Cheng: "Our direction is to let the machine perform deep learning based on the functions of artificial intelligence."

"Deep learning?"

Luo Cheng: "Yes, deep learning! Do you understand deep convolutional neural networks?"

Zhang Xu shook his head. He now feels like a monkey who has just entered human society from the virgin forest. People around him are looking at him with caring eyes for small animals.

Luo Cheng didn't know what he was thinking, and continued to explain: "In terms of face alignment, our model adopts a 3D alignment method, and uses the traditional LBP histogram to texture the image and extract the corresponding features. The extracted The features of the face are processed by SVR (support vector regression) to extract the face and the corresponding six basic points. Do affine transformation according to the six basic points, and then get the corresponding 3 facial key points according to the 67D model, and make triangles based on these points Divide, and finally get the corresponding 3D face."

"The model is in 3D?"

Luo Cheng: "Yes, our machine uses the general cross-entropy loss function in the training process, and uses the general softmax to classify the training data, but it is used in the process from feature vector generation to classification. Some measures of inner product, chi-square distance or siamese are used to calculate the similarity, and the similarity is used for training to generate a classifier."

Wang Hao then said: "The specific alignment process is
The first step: detect the face and the corresponding six basic points

Step [-]: Face after two-dimensional alignment
Step 2: Use Dironi triangulation to divide 67 key points on the [-]D face, and add triangles at the edge to avoid discontinuity

Step 3: Standard 2D face model (required when converting to a 2D plane and comparing with the original 3D picture) (the corresponding average face model generated by the standard [-]D face library USF)

Step 3: The standard 2D face is converted to 2D and the original [-]D is used for residual changes. The black part represents the invisible triangle.For invisible triangles, a symmetrical solution is used
Step 3: Carry out segmental mapping through the 67 reference points generated by the [-]D model to make the face curved and align the face
Step 2: Process the generated [-]D face
Step 3: According to the [-]D face generated by processing. "

Zhang Xu is a little numb, and he is a little dizzy now, why is a punch card machine so high-end?It takes eight steps to successfully clock in. He knows how many steps it takes to put the elephant in the refrigerator. The first step is to open the refrigerator door, and the second step is to put the elephant...

(End of this chapter)