Multi-channel and multi-scale mid-level image representation for scene classification

Abstract

Convolutional neural network (CNN)-based approaches have received state-of-the-art results in scene classification. Features from the output of fully connected (FC) layers express one-dimensional semantic information but lose the detailed information of objects and the spatial information of scene categories. On the contrary, deep convolutional features have been proved to be more suitable for describing an object itself and the spatial relations among objects in an image. In addition, the feature map from each layer is max-pooled within local neighborhoods, which weakens the invariance of global consistency and is unfavorable to scenes with highly complicated variation. To cope with the above issues, an orderless multi-channel mid-level image representation on pre-trained CNN features is proposed to improve the classification performance. The mid-level image representation of two channels from the FC layer and the deep convolutional layer are integrated at multi-scale levels. A sum pooling approach is also employed to aggregate multi-scale mid-level image representation to highlight the importance of the descriptors beneficial for scene classification. Extensive experiments on SUN397 and MIT 67 indoor datasets demonstrate that the proposed method achieves promising classification performance.