研究目的

To propose a novel image classification framework that integrates deep convolutional features and Kernel Extreme Learning Machines (KELMs) for improved performance.

研究成果

In this paper, we propose a novel framework in which deep convolutional features are combined with KELMs for image classification. Densely connected network (DenseNet) is employed as the feature extractor, while a radial basis function kernel ELM instead of linear fully connected layer is adopted as a classifier to discriminate categories of extracted features to promote the image classification performance. Sufficient experiments demonstrate the effectiveness and robustness of the proposed method.

研究不足

1). Redundant training. The CNN and KELM are not trained end to end, which spend much time on selecting features and training the KELM classifier. 2). Large dataset. In the KELM learning step, the amount of selected features should not too large for its high-dimensional matrix calculation. Therefore the proposed framework is hard to apply to large datasets e.g. ImageNet. 3). Hyper-parameter setting. The amount of random hidden nodes in ELM is uncertain, and it significantly influences the final classification results, so we need to constantly debug and test it.