| Image classification is a fundamental issue in the field of pattern recognition,and it has wide applications in the real world.With the popularity of portable digital imaging devices and social media,the image data on the Internet has exploded.Under this circumstance,the study of classification algorithms for large-scale images has become an urgent demand.Due to the size of the data,in the large-scale image classification,the computational efficiency,especially the test efficiency,becomes a critical factor that affects the performance of the classification algorithm.Moreover,these image categories are not completely independent,some categories may have strong inter-category visual similarities,and the main issue is how to use these correlations to improve the performance of large-scale image classification.Hierarchical learning is a method that organizes image categories into tree structures and completing classification according to the structure.It can utilize the correlations between large-scale image categories to build a hierarchical structure,which greatly enhances the testing efficiency and avoid data imbalance.Multi-task learning algorithm can train the inter-related classifiers jointly to boost the performance of each task.One open issue for multi-task learning is how to automatically identify the inter-related tasks and the hierarchical structure has provided a good environment to automatically identify the inter-related tasks for multi-task learning.Based on the above analysis,this thesis proposes a number of hierarchical multi-task algorithms to improve the accuracy of large-scale image classification,while promote the efficiency.The main contributions of this thesis are as follows:1.In dealing with large-scale image classification,many of the existing flat algorithms often partition each image category independently,thus ignoring the correlation between image categories.In addition,the computational cost of flat algorithms is huge.To alleviate the above issue,a novel approach is developed to learn a tree of multi-task sparse metrics hierarchically over a visual tree to achieve a fast solution to large-scale image classification,where an enhanced visual tree is first learned to organize large numbers of image categories hierarchically in a coarse-to-fine fashion.Over the visual tree,a tree of multi-task sparse metrics is learned hierarchically.First,performing multi-task sparse metric learning over the sibling child nodes under the same parent node to explicitly separate their commonlyshared metric from their node-specific metrics.In addition,propagating the node specific metric for the parent node to its sibling child nodes,so that more discriminative metrics can be learned for controlling inter-level error propagation effectively.The experimental results demonstrated that our hierarchical multi-task sparse metric learning algorithm can obtain better performance than the state-of-the-art algorithms on large-scale image classification.2.In multi-task learning,multiple inter-related tasks are learned jointly to achieve better performance.In many cases,when we identify which tasks are related,we can also clearly identify which tasks are unrelated.In the past,researchers often emphasized the related tasks while ignored the unrelated tasks that may provide valuable prior knowledge.To alleviate this issue,a new approach is developed to learn a tree of multi-task metrics hierarchically by leveraging the knowledge about both the related tasks and unrelated tasks.Specifically,an enhanced visual tree is constructed to provide a good environment to automatically identify the related tasks and unrelated tasks.Over the enhanced visual tree,each node is learned as a multi-task metric classifier by exploiting both the related and unrelated tasks.Our experimental results have demonstrated that our hierarchical metric learning algorithm has achieved better results than other state-of-the-art algorithms.3.The general hierarchical learning mainly focuses on improving the testing efficiency,however,when dealing with large-scale image classification,training efficiency tends to be the bottleneck of the algorithm.To alleviate the above issue,a fast hierarchical classification algorithm is developed to integrate visual tree with fast multi-task learning to achieve more effective solution for large-scale image classification.By leveraging the tree structure to separate all the categories hierarchically in a coarse-to-fine fashion,enhanced visual tree can provide a good environment to identify the inter-related tasks,and a multi-task SVM classifier is trained for each parent node to achieve more effective separation of its sibling child nodes.The inter-level visual correlations are utilized to train more discriminative multi-task SVM classifiers and control inter-level error propagation.Moreover,a fast algorithm is developed for learning such multi-task SVM classifiers to improve the training efficiency.The experimental results have demonstrated that our hierarchical classification algorithm has achieved very competitive results on both the classification accuracy and the computational efficiency.4.The traditional scene analysis mainly focuses on outdoor scene recognition rather than indoor scene understanding.However,with the widespread use of depth cameras,we have a new opportunity to handle the indoor scene recognition problem.For this application,we propose a hierarchical multi-task metric multi-kernel learning algorithm to conduct the indoor scene recognition.Specifically,the features of the color images and the depth images are extracted first by using the deep neural networks.Then,utilize multi-task metric learning to learn Mahalanobis metrics which transform features to a correcting feature space for obtaining a better representation.By exploiting hierarchical multi-kernel learning,our algorithm can leverage multiple feature representations to train a more discriminative classifier.The experimental results have demonstrated that our proposed algorithm can lead to better indoor scene recognition. |
PDF Full Text Request