Research Content
Our research covers many aspects of artificial intelligence, which aims to promote people’s lives and improve social productivity through cutting-edge academic research. Our research includes object detection, face recognition, facial expression recognition, motion detection, network multiplexing, etc.
Objection Detection
First of all, we extract the information on the image. And then we will compare the extracted information and the information in the database through the convolution of the neural network. Finally, we mark the object. Our method is characterized by the use of the entire picture as a network of input, and it directly returns the location of bounding box and the category which bounding box belongs to in the output layer.
Face Recognition
By extracting the HOG feature from the face image in the data set, the 31-dimensional fhog operator is extracted for each 8 * 8 pixel size cell. And the data is subjected to positive face and side face training by SVM algorithm to obtain the face detection model. Through the algorithm, the picture of the face detection accuracy can reach 99.4% in the MegaFace data set.
We extract the face data feature and perform critical point regression through a multistage cascade regression tree. And then we use the CLNF (Constrained Local Neural Field) algorithm to achieve the training of the face reference point and get the model used to mark the reference point. Eventually we realized the labeling of the 68 face points on the face, and described the contours of the human face, eyes, eyebrows, nose and mouth.
Expression Recognition
Through the spatial-temporal deep neural network, the context information is extracted in the time domain, and we use the convolution neural network CNN in the spatial domain. And then we get the network convergence, and train the network by a joint loss function, so as to determine the psychological mood of the identified object.
Action Detection
We use the double-stream network method which includes the object detection and motion detection. And then by the corresponding semantic analysis, we ultimately get the results of action recognition. By the action recognition, we can distinguish between different human movements. Obviously, the action detection has far-reaching theoretical research significance and strong practical value.
Network Multiplexing
Network multiplexing is aimed to get one network for multiple tasks, and the network is mainly for algorithm research. Practice has proved that the network can be used in the front-end network layer to achieve structural sharing, thereby reducing network complexity.
Projects
Sitting Detection
Sitting detection is a kind of action detection, its main purpose is to detect people’s different sitting position to get people’s sitting habits and mental state, and by comparing with the standard sitting to achieve the purpose of correcting people sitting position.
Convenience Store Customer Identification
In the convenience store customer identification system, we put the camera next to the different categories of goods to identify the number, gender, dwell time of customers before the goods stay. We can get the degree of love of goods of different customers. This system can better help businesses understand the preferences of customers, and thus adjust the product strategy.
Demo
Objection Detection
Object detection can detect the objects in the picture and video, the technology can effectively improve the efficiency of recognition which has a wide range of applications in various fields.
Face Recognition
Face recognition is a kind of biometrics based on human face feature information. Usually we use a camera to collect a picture or video stream containing many faces, and automatically it detects and tracks the face in the image.
Expression Recognition
Expression recognition refers to the separation of a specific expression from a given static image or dynamic video sequence to determine the psychological mood of the identified object.
Expression Recognition
Action Detection
Action detection will identify the action in the pictures or video, and then help us to determine the type of action and the current psychological activities.