Facial expression bn recognition method based on small data set

Abstract

The invention relates to a facial expression BN recognition method based on a small data set. First, the geometric features and HOG features of the facial expression image are extracted, and the action unit (AU) label sample set is formed through feature fusion and normalization, and then constructed. Facial expression recognition BN structure, and qualitative expert experience into a constraint set between BN conditional probabilities, then introduce the objective function, use convex optimization to solve the maximized form of the objective function, complete the estimation of facial expression recognition BN model parameters, and finally Facial expressions were identified using a joint tree inference algorithm. Compared with the prior art, the facial expression recognition method proposed by the present invention can greatly improve the accuracy of facial expression recognition under the condition of small-scale data sets, and can be widely used in human-computer interaction, security system, medical health diagnosis, video communication and areas such as driver identification and fatigue driving.

Description

technical field [0001] The invention relates to the application field of target recognition of artificial intelligence, image engineering and security systems, and in particular to a facial expression BN recognition method based on a small data set. Background technique [0002] Bayesian network (Bayesian Network, BN) can be expressed as B(G, θ), where G is a directed acyclic graph with n nodes, and n nodes in G represent n random variables. Directed edges represent dependencies between random variables; θ is the conditional probability table associated with each node, denoted as P(X i |P a (X i )). θ quantitatively expresses the node X i with its parent node P a (X i ), the formula (1) is the joint probability distribution of BN: [0003] [0004] Among them, P a (X i ) means X in G i The conditional probability distribution of the parent node set, P(X i |P a (X i )) represents the probability of each value of the variable contained in G given the value of t...

AI Technical Summary

Problems solved by technology

At present, many classic and practical algorithms have been researched and developed in the field of BN's CPT learning, but the implementation and application of these methods are based on large-scale data sets (complete or complete after supplementation), and in practical engineering applications, limited Due to factors such as environment, materials, and time, many experiments cannot be repeated many times, so that the experimental data that can be obtained is small, and the sample size is small. The information that can be expressed in such a small sample data set is not complete enough. The accuracy and reliability of BN parameter learning cannot be guaranteed

Images