基于YOLOX轻量化网络的公路监控场景烟雾识别

Smoke Recognition of Highway Surveillance Scenes Based on YOLOX Light-weight Network

针对公路监控场景下烟雾识别精确度低、传统深层卷积神经网络对计算资源占用量大的问题,基于新型目标识别网络YOLOX构建了公路场景烟雾识别系统。首先收集并标注多公路监控场景下的烟雾图像样本数据,增加Smoke100K部分数据集并整合为公路场景烟雾数据集。同时,搭建并训练多种结构不同层数的YOLOX目标识别网络,利用深度可分离卷积方法对YOLOX网络主体结构进行精简优化,构建nano网络。训练过程中,采用Mosaic数据增强手段对数据集进行扩充,最后导入烟雾图像测试集,通过网络识别精确度、模型大小以及总参数量等指标对4种YOLOX识别网络进行系统测试及对比分析。测试结果表明,YOLOX系列模型均能较好地完成烟雾识别任务,其中nano网络识别精确度达到90.29%,而总参数量仅为m网络的1/25,说明利用数据增强方法提高训练集场景多样性能有效提高网络的泛化能力,在测试场景下提高对烟雾的识别精确度。此外,深度可分离卷积方法能有效压缩模型规模,减少计算量,同时保留网络的特征提取能力。

Aiming at the low accuracy of smoke recognition in highway monitoring scene and high hardware device requirements in traditional convolutional neural network, a smoke recognition system in highway scene was constructed based on the new object detection network YOLOX. Firstly, smoke images with Smoke 100K datasets in multi-road monitoring scenes were collected and labeled. Then4 kinds of YOLOX networks with different layers were built and trained. The depthwise separable convolution was used to simplify and optimize the main structure of YOLOX network, build the nano network. Based on this, Mosaic data enhancement method was used to expand the datasets during training process. Finally, by importing the test datasets, 4 kinds of YOLOX networks with different sizes were systematically tested and compared from the network detection accuracy, model size and total number of parameters. The results show that YOLOX series models can implement the smoke detection task well. The accuracy of nano network is 90.29 % and the total number of parameters is only 1/25 of m network, which means that using data enhancement method to improve the scene diversity of datasets can effectively advance the generalization and recognition accuracy of the network.In addition, depthwise separable convolution can effectively compress the size of the network and reduce the amount of calculation, retaining the feature extraction ability of the network.