CBCNet: A Deep Learning Approach to Urban Images Classification in Urban Computing

Abstract

Urban images classification is an import part of urban computing. It is a challenging task for object detection and classification of urban images due to the high complexity of image contents. i.e., an image may contain buildings, pedestrians, vehicles, roads, etc. In this paper, a novel convolutional neural network, named Complex Background Classification Network (CBCNet), is proposed for classifying the urban images. This network is unlike existing AlexNet, ResNet, etc. It uses a multilayer perceptron convolutional layer to extract more representative features of complex urban images, instead of using a linear convolutional layer, and integrates back-propagation network to optimize object parameters. We also build a standard dataset of urban images containing eight categories, contrast experiments prove that the dataset is rational and feasibility. Experiment results obtained on two benchmark datasets demonstrate that classification accuracy and computation of CBCNet outperform those by the previous state-of-the-art items of AlexNet, VGGNet16 and ResNet50.