针对复杂背景的城管案件图像分类方法

Image Classification Method of Urban Management Cases Based on Complex Background

为解决智慧城市管理系统中复杂背景案件图片分类问题,提出了一个卷积神经网络模型。不像其它深度学习网络模型,所提方法在卷积网络中使用多层感知卷积层替代传统线性卷积层来提取城管图像特征,然后再使用一个评价网络根据所提取的特征对图像中所含关键物体进行评分,从而得出分类结果。基于实际运行的城市管理系统整理了一份包含8类案件、4996个注释对象共计4119张图片的城市管理案件图像的标准数据集来验证所提方法。实验结果表明,所提方法在分类精度及速度上都优于AlexNet,VGG16Net,和ResNet分类方法。

In this paper,a novel deep convolution neural network is proposed to classify the images with complex background in the intelligent city management system.Unlike some existing deep neural networks,the algorithm used a multilayer perceptron instead of the commonly used linear convolution layer in convolution network to extract different characteristics of the urban management images.Then an evaluation network integrated the obtained characteristics to achieve the scores of all the urban management objects.To evaluate the algorithm,we also created a standard image database of city management cases that contains 8 categories and 4996 comment objects in total 4119 pictures.The comparison results demonstrate that the classification accuracy and computation speed of the algorithm exceed those of AlexNet,VGG16 Net,and ResNet.