Application of Noise Detection Using Confidence Learning in Lightweight Expression Recognition System

Facial expression is an important carrier to reflect psychological emotion, and the lightweight expression recognition system with small-scale and high transportability is the basis of emotional interaction technology of intelligent robots. With the rapid development of deep learning, fine-grained expression classification based on the convolutional neural network has strong data-driven properties, and the quality of data has an important impact on the performance of the model. To solve the problem that the model has a strong dependence on the training dataset and weak generalization performance in real environments in a lightweight expression recognition system, an application method of confidence learning is proposed. The method modifies self-confidence and introduces two hyper-parameters to adjust the noise of the facial expression datasets. A lightweight model structure combining a deep separation convolution network and attention mechanism is adopted for noise detection and expression recognition. The effectiveness of dynamic noise detection is verified on datasets with different noise ratios. Optimization and model training is carried out on four public expression datasets, and the accuracy is improved by 4.41% on average in multiple test sample sets. A lightweight expression recognition system is developed, and the accuracy is significantly improved, which verifies the effectiveness of the application method.