时空特征学习算法及其应用研究集成电路工程专业论文.docxVIP

时空特征学习算法及其应用研究Abstract 时空特征学习算法及其应用研究 Abstract Most hand—designed features work well only after meticulously designing and selecting．Compared with hand-designed features，feature learning methods directly learn features from data．It eases the hard of designing or selecting features and the dependence on domain knowledge．In recent years，feature learning methods have made a great success in many applications of computer vision and pattern recognition． This paper researches feature learning methods based on convolutional neural networks(CNN)，independent subspace analysis(ISA)and independent component analysis(ICA)．These methods are applied to tasks like facial landmark detection and video classification．We mainly study feature learning methods based on ISA and ICA． And further research multi-layer architecture．The major work of this thesis is shown below： CNN and its application to facial landmark detection are explored．A detailed comparison between many CNNs with different network structures is made．The experiment is carried out based on large scale dataset．The experiment results show that the detection accuracy of Our adjusted network is almost equivalent to that of popular CNN with multi-task learning，but the computation cost is much lower and the necessary labels are less in Our adjusted network． A spatiotemporal feature learning method based on ISA and its application on facial expression recognition are explored．A detailed analysis of how different subspace size affects the robust of learned features for misaligned facial expression recognition is given． Furthermore，a spatiotemporal feature learning method combined with multi-layer ISA and spatiotemporally pooling is proposed．The proposed method achieves excellent performance on facial expression recognition based on MMI database． A spatiotemporal feature learning method based on ICA and its application on fire detection are explored．Furthermore，a spatiotemporal feature learning method with combination of ICA le