Research On The Deep Models For Slow Feature Analysis

Image and video are two most intuitive medias through which human being perceive the outside world.Visual information processing has always been an attractive research topic,among which feature learning is one of the most fundamental problems.Feature learning aims at automatically obtaining a robust yet discriminative representation of visual data,and it has significant impacts on image classification,identification and retrieval.For example,along with the explosive growth of on-line images and videos,there have been an increasing number of illegal copies,and the unauthorized circulation of these copies could arise intellectual property disputes.Copy detection is an effective measure for tackling this problem,and invariant feature extraction is at the core of copy detection.Taking image and video copy detection as application scenarios,this thesis studies the design of feature learning based robust hashing.A review of existing feature learning models is presented in Chapter 2,including sparse coding,deep learning,and slow feature analysis,and these models are the theoretic basis of the research works in following chapters.For image feature learning,we propose a group sparse coding based approach,and the corresponding dictionary learning and coding algorithms are designed to capture the spatially slow-varying visual features.Inspired by the hierarchical structure of human visual system,the proposed work uses a multi-layer group sparse coding model to learn slow-varying features at multiple scales.Experimental results demonstrate that the image hashing algorithm based on this feature learning method shows satisfactory robustness and discriminability,and its performance outperforms state-of-the-arts.This thesis also proposes a spatio-temporal deep neural network for learning invariant video features and applied it in robust video hashing.The proposed first builds a deep neural network to extract features on frame level by greedily training denoising autoencoders(DAE).Consequently,a long short-term memory network(LSTM)is trained to model temporal variation of spatial features,and our training algorithm is designed from the prospective of slow-feature analysis.Experiment results confirm that the proposed algorithm shows high accuracy in video copy detection with an F1 score of 0.995.In summary,based on the theory of spatial and temporal slow feature analysis,with sparse coding and deep neural network as computational models,this thesis proposed two self-learning methods for extracting invariant image and video descriptors,and the research works of this thesis could benefit the applications of content identification and retrieval.