Structural Learning for Large Scale Image Classification
Analytics
48 views ◎17 downloads ⇓
Abstract
To leverage large-scale collaboratively-tagged (loosely-tagged) images for training a large number of classifiers to support large-scale image classification, we need to develop new frameworks to deal with the following issues: (1) spam tags, i.e., tags are not relevant to the semantic of the images; (2) loose object tags, i.e., multiple object tags are loosely given at the image level without their locations in the images; (3) missing object tags, i.e. some object tags are missed due to incomplete tagging; (4) inter-related object classes, i.e., some object classes are visually correlated and their classifiers need to be trained jointly instead of independently; (5) large scale object classes, which requires to limit the computational time complexity for classifier training algorithms as well as the storage spaces for intermediate results.To deal with these issues, we propose a structural learning framework which consists of the following key components: (1) cluster-based junk image filtering to address the issue of spam tags; (2) automatic tag-instance alignment to address the issue of loose object tags; (3) automatic missing object tag prediction; (4) object correlation network for inter-class visual correlation characterization to address the issue of missing tags; (5) large-scale structural learning with object correlation network for enhancing the discrimination power of object classifiers.To obtain enough numbers of labeled training images, our proposed framework leverages the abundant web images and their social tags. To make those web images usable, tag cleansing has to be done to neutralize the noise from user tagging preferences, in particularly junk tags, loose tags and missing tags. Then a discriminative learning algorithm is developed to train a large number of inter-related classifiers for achieving large-scale image classification, e.g., learning a large number of classifiers for categorizing large-scale images into a large number of inter-related object classes and image concepts. A visual concept network is first constructed for organizing enumorus object classes and image concepts according to their inter-concept visual correlations. The visual concept network is further used to:(a) identify inter-related learning tasks for classifier training;(b) determine groups of visually-similar object classes and image concepts; and(c) estimate the learning complexity for classifier training.A large-scale discriminative learning algorithm is developed for supporting multi-class classifier training and achieving accurate inter-group discrimination and effective intra-group separation. Our discriminative learning algorithm can significantly enhance the discrimination power of the classifiers and dramatically reduce the computational cost for large-scale classifier training.