Machine vision identification model for reconstructed and expanded brick-concrete houses with hidden dangers
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 7 114 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,F1-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 0.945 9 and a recall rate of 0.974 2 on the test set,showcasing a notable improvement of 0.014 8 and 0.044 4,respectively,compared to the best-performing model trained without RICAP augmentation.What's more,KMeans++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.
safety engineeringreconstructed and expanded housemachine visionRandom Image Cropping and Patching(RICAP)transfer learning
万方数据
维普
