Key Technology Research On Hierarchical Classification

The special architecture of hierarchical classification makes it suitable for distributed computing,which is important and beneficial for large-scale image classification.Besides,for a given tree node,its classifier focuses on discriminating a small number of sibling child nodes,thus the complexity for classifier training can be reduced significantly and the problem of huge sample imbalance can be under control,while the discrimination power of the node classifier can be improved dramatically.Hierarchical classification consists of three important parts,the construction of category tree,classifier training and feature learning.Most existing works on hierarchical classification focus on the classifier design and category tree,while they have ignored the poor discrimination power of features.To solve this problem,we mainly focus on feature learning.The major contributions and innovations are summarized as follows:1)We propose a deep multi-task learning algorithm(DMTL)to combine hierarchical classification with deep convolutional neural networks,which makes the category hierarchy embedding into deep networks.Recently,deep learning methods have shown powerful ability in feature learning,and achieved much better performances than hand-engineering features in many tasks.Based on the structure of category tree,we leverage a deep multi-path convolutional neural networks and the inter-task relatedness to learn more discriminative group-specific deep representations,and our deep multi-task learning algorithm can train more discriminative node classifiers for distinguishing the visually-similar atomic object classes effectively.The experimental results show that the DMTL algorithm has improved the classification performance.2)We propose Laplacian pyramid auto-encoders and Jigsaw puzzles auto-encoders,which leverage scale information and spatial information for unsupervised feature learning.Laplacian pyramid auto-encoder is a straightforward modification of the deep convolutional auto-encoder architecture.The method we propose here uses multiple encoding-decoding sub-networks within a Laplacian pyramid framework to reconstruct the original image and the low pass filtered images.Jigsaw puzzles auto-encoders leverage convolutional auto-encoders or siamese-ennead auto-encoders to solve Jigsaw puzzles,which make the encoding networks succeed in capturing the spatial information.3)In order to deal with multi-modal data,we apply Canonical Correlation Analysis algorithm and coupled dictionary learning in multi-modal feature learning.Canonical Correlation Analysis algorithm focuses on the common feature subspace learning.Coupled dictionary learning based on structured sparseness can learn a common latent space across different modalities,which capture the semantic relationships between different modalties and allows each latent factor to be associated with a subset of modalities.4)In the real world,human tends to select one special kind of visual features to group the objects at different semantic levels.Inspired by this phenomenon,we select one special visual modality by the evaluation of each modality's discriminative ability at each divisive step in the top-down hierarchical clustering procedure,which results in a modality-sensitive algorithm.Experimental results show that this algorithm works well,and the keypoint is to define the selective standard in an unsupervised way.