基于深度多任务学习的微表情识别研究

微表情研究在国家安全、司法审讯、机场安检、案件侦破以及监狱管理等发挥着巨大的作用，日益被广泛应用在医学领域、人工智能领域、情绪智能研究领域、教育领域、交通领域等。然而微表情的特征易受到噪声等影响，并且微表情数据库中训练样本较少给微表情识别带来巨大的挑战。本课题拟开展基于深度多任务学习的微表情识别研究。针对深度学习中的共享层无法确定，网络参数过多的问题，开展共享层自适应学习研究，构建张量迹范数正则化约束下的多任务学习模型；为选取有代表性和判别力的特征，提高小样本下微表情识别性能，开展稀疏层次多任务反卷积网络研究，将视觉词典生成和分类器优化组合；针对微表情数据库中缺少训练样本，识别精度不高问题，开展基于分类和特征学习的深度多任务学习研究，将微表情和宏表情有机融合；进而，通过多生物识别下的多任务深度网络模型研究，设计强分类器，充分利用生物识别信息，解决多源多任务学习的问题，提升微表情的识别性能。

Micro-expression research plays an important role in national security, judicial trial, Airport security, case detection and prison management. It is increasingly widely used in the fields of medicine, artificial intelligence, emotional intelligence, education, transportation and so on. However, the characteristics of micro-expressions are easily affected by noise and other factors. There are fewer training samples in existing micro-expressions databases results in poor performance of micro-expressions recognition, which brings great challenges to micro-expressions recognition.This topic intends to carry out micro-expression recognition research based on deep multi-task learning. In order to solve the problems of uncertain sharing layer and excessive network parameters in deep learning, adaptive learning of sharing layer is studied, and a multi-task learning model under the constraints of tensor trace norm regularization is constructed. In order to select representative and discriminant features and improve micro-expression recognition performance under small samples, sparse hierarchical multi-task deconvolution network is studied to combine the visual dictionary and classifier. Aiming at the lack of training samples and low recognition accuracy in micro-expression database, deep multi-task learning based on classification and feature learning is carried out to integrate micro-expression and macro-expression organically; furthermore, through the multi-task deep network model research under multi-biometric recognition, a strong classifier is designed to make full use of biometric information to solve multi-purpose and multi-task problems, and improve the recognition performance of micro-expressions.