WCSE

GraftNet: An Efficient and Flexible Multi-label Image Classification and Its Application

Abstract— A multi-label network with an adaptive hierarchical structure is presented in this paper. The proposed network is efficient and flexible; we call it GraftNet in abbreviation. The network is in a tree-like fashion which has a trunk and several branches. The trunk is pre-trained with a dynamic graph for generic feature extraction and branches separately trained on sub-datasets with single label to improve accuracy and efficiency. We employ efficient neural architecture search (ENAS) for the branches part which outperforms manually designed architectures on image classification. The proposed network could avoid the inherent problem of “catastrophic forgetting” problem in deep neural networks and the addition of new image classes to a network could be easier than retraining the whole network again with all the classes. The employment of ENAS helps our network to achieve high performance. The proposed hierarchical network is also suitable for imbalanced dataset. Compared against fine-tuning a deep network, the proposed network achieves significant reduction of training and testing effort. Experimental results show that it has good performance on our human attributes recognition task.

Index Terms— Convolutional Neural Networks· Deep Learning· Incremental Learning· Transfer Learning · NAS.

Chunhua Jia, Shuai Zhu, Wenhai Yi, Yu Wu, Leilei Wu, WeiweiCai
Shanghai Elevator Media Information Co., Ltd.,CHINA

[Download]

Cite: Chunhua Jia, Shuai Zhu, Wenhai Yi, Yu Wu, Leilei Wu, WeiweiCai , "GraftNet: An Efficient and Flexible Multi-label Image Classification and Its Application," 2021 The 11th International Workshop on Computer Science and Engineering (WCSE 2021), pp. 70-78, Shanghai, China, June 19-21, 2021.