TY - GEN
T1 - Ensemble of binary tree structured deep convolutional network for image classification
AU - Lee, Ji Eun
AU - Kang, Min Joo
AU - Kang, Je Won
N1 - Funding Information:
This work was supported by Institute for Information and communications Technology Promotion(IITP) grant funded by
Publisher Copyright:
© 2017 IEEE.
PY - 2018/2/5
Y1 - 2018/2/5
N2 - In this paper, we propose an ensemble of tree- structured learning architecture to improve the discriminative capability of deep convolutional neural network (DCNN) for image classification. In the proposed technique, the path from the root node to a leaf node represents a classification rule. Thus, to maximize the classification accuracy, each internal node needs to make an optimal binary decision to the left or the right child node. To this aim, we develop a tree-CNN as a randomized tree to embed a DCNN into each internal node and train the model to determine the best traversing path to predict a class. Classification of some images with similar statistical properties yet belonging to different classes are difficult with the conventional DCNN architecture. Thus, to resolve the problem, we use a coarse-to-fine approach where subsequent networks in children nodes are hierarchically and randomly organized to discriminate smaller sets of classes than those in a parent node. The results from all the individual tree-CNNs are ensembled to make the final decision in classification. The proposed technique is implemented with the state-of-the-art deep network model, i.e., Wide Residual Network DCNN model [19], and is demonstrated with experimental results to outperform the classification performance over the anchor.
AB - In this paper, we propose an ensemble of tree- structured learning architecture to improve the discriminative capability of deep convolutional neural network (DCNN) for image classification. In the proposed technique, the path from the root node to a leaf node represents a classification rule. Thus, to maximize the classification accuracy, each internal node needs to make an optimal binary decision to the left or the right child node. To this aim, we develop a tree-CNN as a randomized tree to embed a DCNN into each internal node and train the model to determine the best traversing path to predict a class. Classification of some images with similar statistical properties yet belonging to different classes are difficult with the conventional DCNN architecture. Thus, to resolve the problem, we use a coarse-to-fine approach where subsequent networks in children nodes are hierarchically and randomly organized to discriminate smaller sets of classes than those in a parent node. The results from all the individual tree-CNNs are ensembled to make the final decision in classification. The proposed technique is implemented with the state-of-the-art deep network model, i.e., Wide Residual Network DCNN model [19], and is demonstrated with experimental results to outperform the classification performance over the anchor.
UR - http://www.scopus.com/inward/record.url?scp=85050471277&partnerID=8YFLogxK
U2 - 10.1109/APSIPA.2017.8282260
DO - 10.1109/APSIPA.2017.8282260
M3 - Conference contribution
AN - SCOPUS:85050471277
T3 - Proceedings - 9th Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2017
SP - 1448
EP - 1451
BT - Proceedings - 9th Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2017
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 9th Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2017
Y2 - 12 December 2017 through 15 December 2017
ER -