基于卷积神经网络的灾难场景图像分类

Image Classification of Disaster Scenes Based on Convolutional Neural Network

为了对灾难场景图像进行快速分析和识别,提出了一种基于多分辨率卷积神经网络和残差注意力机制(attention module)相结合的图像分类模型。首先,对灾难场景数据集进行预处理,在相同类型的条件下将其随机划分为训练集和测试集。基于改进的卷积神经网络(convolutional neural network,CNN),提取训练集的图像特征。使用训练集图片的特征进行训练,并且在测试集图片上实现分类测试。选取DenseNet、Xception和MobileNetV2三种模型,以灾难场景数据集(Disaster_Data_Scenes)为数据集进行实验验证。结果表明:改进的Xception和MobileNetV2网络在灾难场景数据集上进行的图像分类实验测试,比原网络精度分别提升了4.56%和3.04%。其中改进的DenseNet网络比原网络模型精度分别提升9.13%、17.88%和10.27%。可见改进的卷积神经网络模型的分类精度得到有效提高。

In order to quickly analyze and recognize disaster scene images,an image classification model based on the combination of multi-resolution convolutional neural network and the attention module was proposed.Firstly,the disaster scenario data set was preprocessed and randomly divided into training set and test set under the same type of conditions.The image characteristics of the training set were extracted by an improved convolutional neural network(CNN).The characteristics of the training set images were used for training,and the classification test was implemented on the test set images.The three models of DenseNet,Xception and MobileNetV2 were selected and the disaster scene data set(Disaster_Data_Scenes)was used for experimental verification.The results show that the improved Xception and MobileNetV2 network improves the accuracy of image classification by 4.56%and 3.04%,respectively,compared with the original network.Among them,the accuracy of the improved DenseNet network is 9.13%,17.88%and 10.27%higher than that of the original network model.It is concluded that the classification accuracy of the improved convolutional neural network model is effectively improved.