DELTA: A deep dual-stream network for multi-label image classification

作者：

Highlights：

• We propose an end-to-end deep model with dual streams, which is able to effectively extract and make use of the global image priors and the local image features.

• We incorporate a spatial pyramid convolutional transfer layer into the deep model.

• We utilize multi-instance pooling layer to effectively aggregate the information contained in the feature maps.

• Extensive experiments demonstrate that DELTA outperforms state-of-the-art deep multi-label classification methods.

摘要

•We propose an end-to-end deep model with dual streams, which is able to effectively extract and make use of the global image priors and the local image features.•We incorporate a spatial pyramid convolutional transfer layer into the deep model.•We utilize multi-instance pooling layer to effectively aggregate the information contained in the feature maps.•Extensive experiments demonstrate that DELTA outperforms state-of-the-art deep multi-label classification methods.

论文关键词：Deep neural network,Dual-stream network,Multi-label image classification,Multi-instance learning

论文评审过程：Received 9 October 2018, Revised 24 February 2019, Accepted 10 March 2019, Available online 11 March 2019, Version of Record 13 March 2019.

论文官网地址：https://doi.org/10.1016/j.patcog.2019.03.006