Abstract: In this paper, an improved deep Convolution Neural Network（CNN） algorithm is proposed to classify the urban management cases’ images in city management system. CNN can extract the image features automatically. The procedure can be divided into three main stages. Initially, ZCA-whitening is used to reduce the correlation of the images. Then, an eight-layer CNN model is set up to classify the images processed by ZCA-whitening. In convolution layer, ReLU is employed to speed up the training phase, and dropout is used to prevent overfitting in pooling layer. Finally, BP algorithm is introduced to improve the robustness of algorithm in fine-tuning phase. The method achieves about 97.5% accuracy and the F1-Score is 0.98 on two classes’ images of traffic and environment. The performance exceeds the LSVM, SAE and the usual CNN algorithm. Also, it tests on four classes images of electric-bicycles, rubbish, cars and dustbins, the accuracy is 90.5%, F1-Score is 0.91, and the performance still exceeds the LSVM, SAE and the usual CNN algorithm.

Key words: urban management, image classification, Convolution Neural Network（CNN）, Zero-phase Component Analysis（ZCA）-whitening, dropout, ReLU

摘要： 以智慧城市管理系统中上报的案件图像为研究对象，利用卷积神经网络能够自行学习图像特征的优势，提出一种改进的深层卷积神经网络算法，并利用该算法对智慧城市管理系统（下简称“智慧城管”）的案件图像进行快速精确分类，从而完成城市管理系统中案件的自动分类。采用ZCA白化处理降低图像数据特征之间的相关性；搭建八层卷积神经网络对白化后的图像进行分类，并在卷积层采用线性纠正单元（ReLU）加速训练过程，在pooling层利用dropout技术防止算法过拟合；在网络精调阶段采用BP（Back Propagation）算法进行优化，提高算法的鲁棒性。基于上述方法对道路交通类和市容环境类两类案件图像进行二分类实验，平均精度达到97.5%，F1-Score达到0.98，性能超过了LSVM、SAE以及传统的CNN等方法；同时该方法又对电动车乱摆放类、乱扔垃圾类、机动车违章停放类、垃圾桶周围脏乱类共四类案件进行四分类实验，平均精度为90.5%，F1-Score为0.91，性能依然超过了LSVM、SAE以及传统的CNN等方法。

关键词: 智慧城管, 图像分类, 卷积神经网络, 零相位分量分析（ZCA）白化, dropout, ReLU