Convolution neural network and conditional random field-based face detection method

Abstract

The present invention discloses a convolution neural network and conditional random field-based face detection method. With the method adopted, the accuracy of a face frame determined by a network output layer can be improved. The method includes the following steps that: a convolution neural network is trained, so that a classifier for identifying faces and non-face objects can be obtained, and sliding window face detection is carried out on an input image, so that windows containing faces can be obtained; all detection windows containing the same face are marked, confidence scores corresponding to the windows are adopted as the random variables of a conditional random field (CRF), and correlation relationships between the windows are calculated through a CRF model, and the windows are selected according to the degree of the closeness of the correlation relationships; and windows of the same scale and different scales are combined according to the size of area overlapping, horizontal distance and the size of vertical distance overlapping, and therefore, a final face frame can be obtained.

Description

technical field [0001] The invention belongs to the field of computer machine learning, and more specifically relates to a face detection method based on a convolutional neural network and a conditional random field. Background technique [0002] Face Detection technology is an important branch of machine learning. Its core idea is to first train a large number of face data samples to obtain a two-class classifier of a face, and then give a picture as input , through a certain strategy to select any area in the picture for discrimination, frame the area that is the face, until all areas of the picture are discriminated. [0003] At present, face detection methods mainly include cascade classifier (Cascade), regional feature detection method DPM, and convolutional neural network detection method, among which convolutional neural network detection method is the mainstream method widely used in recent years. [0004] The basic principle of convolutional neural network is to tr...

AI Technical Summary

Problems solved by technology

These algorithms have good performance for frontal faces, but they are still not enough to deal with partial faces, expressive faces and occluded faces, and the accuracy rate is low when dealing with different angles, different expressions and occluded faces.

Images