Sparse decomposition of convolutional features for scene recognition
Lin Xie, Feifei Lee, Yan Yan, Qiu Chen · 2017
Scene recognition is an important and challenging problem in the field of computer vision owing to the variations in the same class and the similarities between different classes. This paper presents a novel approach that learns a reasonable dictionary from convolutional features to effectively describe the distinctive and shared properties in scene images. Substantial convolution operations in Deep Convolutional Neural Networks (DCNN) make the output features have redundant and irrelevant information for scene recognition. Sparse Decomposition of Convolutional Features (SDCF) aims to extract the effective components for scene recognition. Convolutional features are apt to capture the holistic appearance and ignore the partial details because the DCNN is trained by the whole images. So we propose to take advantage of the convolutional features of densely sampled patches to compensate for the lack of the local information. The experimental result shows that our approach outperforms the state-of-the-art results on the benchmark.