Compressed Holistic Convolutional Neural Network-based Descriptors for Scene Recognition

Deep convolutional neural networks (CNN) have recently been widely used in many computer vision and pattern recognition applications. With the help of high-level image description features provided by CNN, the deep architecture models perform significantly better than state-of-the-art solutions that use traditional hand-crafted features. In this paper, we concentrate on the scene recognition problem especially for changing environments, such as view angle changes, illumination variations, occlusion, different weather conditions and seasons. We propose a new scene recognition system using the deep residual convolutional neural network (ResNet) as the image feature extractor. The initial feature vectors are chosen from specific layers of the network and after a series of post-processes, we can obtain the final image descriptor vectors for scene similarity measurement. The performance of our proposed methods is evaluated on four popular open datasets by comparing it with the classic FabMap method and some other deep learning-based methods.