Rapid face detection and recognition method for DSP platform

Abstract

The invention discloses a rapid face detection and recognition method for a DSP platform. In order to solve the problem, the whole face detection and recognition process is roughly as follows: a cascaded P-R-O network (namely MTCNN) is responsible for simultaneously detecting a face and key feature points in an image, a source image is cut into the size of 3 * 299 * 299 according to the size of a finally output face frame, and the cut source image is used as the input of a recognition network FACENET; and whether the human faces are similar or not is judged by comparing the distance between the corresponding 512-dimensional normalized feature vectors of the image to be detected and other images in the Euclidean space.

Description

Technical field: [0001] The invention relates to the fields and directions of artificial intelligence, deep learning neural network, image processing, computer vision, edge computing, face detection and recognition, etc. Background technique: [0002] In various computer vision tasks, such as scene understanding, anomaly detection, and semantic segmentation, convolutional neural network (CNN) has achieved better performance. big boost. In order to meet the low-latency, high-precision, and real-time requirements of CNN forward propagation performance in practical applications, deploying CNN to edge parallel computing devices such as GPUs, DSPs, FPGAs, and ASICs has become a current research hotspot in the industry. [0003] Compared with other biometrics (fingerprint, voice, gait, etc.), the face has more significant individual differences and uniqueness, and it also has certain advantages in collection. It is important to use face detection and recognition technology for au...

AI Technical Summary

Problems solved by technology

[0010] The above prior art methods have the following disadvantages: the method of deleting parameters in units of convolution kernels is too rough [1] , in a 3D convolution kernel, there are few parameters that have a greater impact on prediction accuracy, and this pruning technique is often inefficient

Patent [4][5][6][7] is only the application of neural network, and does not optimize the network structure and calculation process, often need to compromise and weigh speed and accuracy

Images