融合双注意力机制EfficientNet的沥青路面状态分类方法

针对现有EfficientNet模型应用于沥青路面状态分类时，卷积操作易导致高层特征信息丢失问题，在现有EfficientNet模型的深层结构中引入一种双注意力机制，包含通道注意力模块和位置注意力模块，借助Sigmoid线性单元（Sigmoid linear unit，SiLU）激活函数和余弦学习率衰减策略，提出一种融合双注意力机制EfficientNet （Dual attention network based on EfficientNet，DAEfficientNet）的沥青路面状态分类方法。首先，建立不同天气下5种沥青路面共5 938张图像作为数据集，积雪样本来自开源数据集（Canadian adverse driving conditions dataset，CADCD）。然后，对所提出模型进行训练，并得到沥青路面图像分类结果。最后，利用准确率（Accuracy）、精确率（Precision）、召回率（Recall）、F1 score和特异度（Specificity），将所提出模型与其他现有卷积神经网络模型进行分类效果对比分析。试验结果表明：所提出模型优于其他对比模型，能准确、有效地对不同天气下的沥青路面状态进行分类。

Dual Attention Network for the Classification of Road Surface Conditions Based on EfficientNet

Given the drawback of information loss of high-level features caused by convolution operations when the existing EfficientNet is applied to classify asphaltroad surfaces, a novel dual attention mechanism combined two types of attention modules, named channel attention module and position attention module respectively, is introduced to the existing EfficientNet, and the dual attention network based on EfficientNet (DAEfficientNet) is proposed using sigmoid linear unit (SiLU) activation functions and a cosine learning rate decay technique. First, a dataset including 5,938 images of five types of asphalt road surfaces under various weather conditions is constructed. The snow image samples of asphalt road surfaces are from the open-source dataset named Canadian adverse driving conditions dataset (CADCD). Second, the proposed model is trained and the image classification results are produced. Finally, the accuracy, precision, recall, F1 score, and specificity of the analyzed models are calculated to compare the classification performance between the proposed model and the others previous convolutional neural network models. The experimental results show that the proposed method outperforms the others completing methods and achieves higher accuracy and stronger robustness in the task of classification of the five types of road surface images.