改扩建砖混隐患房屋机器视觉辨识模型

Machine vision identification model for reconstructed and expanded brick-concrete houses with hidden dangers

改扩建砖混房屋存量巨大,且易发生坍塌事故。仅靠人工方式辨识隐患房屋较为低效,尝试引入机器视觉模型更智能地辨识隐患房屋。首先,收集已标注安全现状的相关房屋图像7 114幅,经一定处理后按6∶2∶2的比例划分训练集、验证集与测试集;之后,选用ConvNeXt、Swin Transformer、ResNeXt、DenseNet和MobileNet五种视觉模型,以迁移学习的方法,进行适应本识别任务的微调训练,并与引入“随机裁剪拼接(Random Image Cropping and Patching, RICAP)”图像数据增强方法的训练结果对比;最后,综合运用多种指标评价各个模型表现。结果表明:引入RICAP方法训练的ConvNeXt模型,在测试集上取得0.945 9的准确率、0.974 2的召回率,比未使用该方法训练的表现最优的模型分别提升0.014 8、0.044 4,可更准确地辨识隐患房屋。

There is a huge stock of reconstructed and expanded brick-concrete houses,which are prone to collapse accidents.Traditionally,the identification of these hidden danger houses relies on manual inspections,which is inefficient.This study attempts to leverage machine vision models to intelligently identify houses with hidden dangers.In the first place,a dataset,consisting of 7114 images of reconstructed and expanded brick-concrete houses from a province in southern China,was collected for the research effort.Each image was labeled with house safety status.These images underwent preprocessing before being divided into three subsets:training set,validation set,and test set,distributed in a 6:2:2 ratio.Secondly,the study employed the approach of transfer learning,fine-tuning five diverse machine vision models:ConvNeXt,Swin Transformer,ResNeXt,DenseNet,and MobileNet.The data augmentation method,Random Image Cropping and Patching(RICAP),was applied during training to further enhance model robustness and generalization.RICAP is a method that randomly crops four images and patches them together to generate a new training image.It also mixes the class labels of the four images,resulting in a benefit similar to label smoothing.Finally,a comprehensive evaluation of each network s performance was conducted using multiple metrics.These metrics include confusion matrix,accuracy rate,recall rate,precision rate,specificity,F 1-score,ROC curve,etc.The results demonstrate the superiority of the ConvNeXt model when trained with the RICAP augmentation method.This model achieved an impressive accuracy rate of 0.9459 and a recall rate of 0.9742 on the test set,showcasing a notable improvement of 0.0148 and 0.0444,respectively,compared to the best-performing model trained without RICAP augmentation.What s more,K Means++algorithm was applied to cluster the images of hazardous houses,further validating the ConvNeXt(RICAP)model,which achieved the highest accuracy in this research,also demonstrated a capability to differentiate between specific types of hazards of houses and assess their severity.