Source-linked AI summary
Deep Learning for Detecting Building Defects Using Convolutional Neural Networks
Husein Perez, Joseph H. M. Tah, Amir Mosavi
TL;DR
Building condition assessment is time-consuming, expensive, and hazardous, creating a need for faster image-based inspection. The paper evaluates a transfer-learning CNN approach using VGG-16 to detect and localise dampness-related defects, reporting strong classification performance while identifying real-world limitations.
Problem
Building condition assessment relies on costly, laborious inspections, while deep learning also faces challenges from limited and distribution-mismatched training data.
Method
The paper fine-tunes a VGG-16 network pretrained on ImageNet to classify building images and integrates CAM for defect localisation.
Results
98.86% validation accuracy was recorded after 50 epochs, and the model accurately detected and localised mould, stains, and paint deterioration.
Takeaways & Limitations
The approach demonstrates automated multi-class detection and localisation of dampness-related building defects using relatively small image datasets.
Takeaways & Limitations
The study excludes multiple defect types in one image, invisible defects, and extreme lighting or orientation conditions.
Abstract
from arXiv · showhide
Clients are increasingly looking for fast and effective means to quickly and frequently survey and communicate the condition of their buildings so that essential repairs and maintenance work can be done in a proactive and timely manner before it becomes too dangerous and expensive. Traditional methods for this type of work commonly comprise of engaging building surveyors to undertake a condition assessment which involves a lengthy site inspection to produce a systematic recording of the physical condition of the building elements, including cost estimates of immediate and projected long-term costs of renewal, repair and maintenance of the building. Current asset condition assessment procedures are extensively time consuming, laborious, and expensive and pose health and safety threats to surveyors, particularly at height and roof levels which are difficult to access. This paper aims at evaluating the application of convolutional neural networks (CNN) towards an automated detection and localisation of key building defects, e.g., mould, deterioration, and stain, from images. The proposed model is based on pre-trained CNN classifier of VGG-16 (later compaired with ResNet-50, and Inception models), with class activation mapping (CAM) for object localisation. The challenges and limitations of the model in real-life applications have been identified. The proposed model has proven to be robust and able to accurately detect and localise building defects. The approach is being developed with the potential to scale-up and further advance to support automated detection of defects and deterioration of buildings in real-time using mobile devices and drones.
1. Introduction
Building condition assessment is slow, costly, and potentially hazardous, motivating image-based automation. This study investigates CNNs with transfer learning and CAM to detect and localise dampness-related defects.
- Traditional assessments require lengthy site inspections, systematic recording, reporting, and estimates of immediate and projected repair costs.
- Image analysis could replace manual inspections, but performance depends strongly on image quality under varying light, shadow, and noise.
- The research investigates CNNs for automated detection and localisation of key dampness-related building defects from images.
- The initial study excludes simultaneous multiple-defect images, invisible defects, and extreme lighting or orientation conditions.These limitations are identified as targets for future work toward fully automated detection.
- The proposed approach uses transfer learning with VGG-16 for feature extraction and classification, with CAM supporting defect localisation.
2. Dampness in Buildings
Dampness contributes to diverse building deterioration and can affect building value, materials, indoor conditions, and occupant health. The paper focuses on mould, stains, and paint deterioration as common interrelated dampness-related defects.
- Dampness is moisture present in sufficient quantity to be visible or touchable, or to cause deterioration in decorations and building fabric.
- Excessive dampness can trigger corrosion, fungal attack, efflorescence, sulphate attack, carbonation, material degradation, and mould growth.
- Condensation, rain penetration, and rising damp account for a high proportion of dampness problems.
- Rain penetration commonly affects roofs and wind-exposed walls through openings, faulty design or workmanship, structural movement, material decay, repairs, or insufficient maintenance.
- Rising damp moves upward through capillary action from damp soil into lower wall sections and is common in older properties with damaged or absent damp-proof courses.
- The study focuses on moulds, stains, and paint deterioration, which are described as common interrelated defects arising from dampness.
3. Convolutional Neural Networks (ConvNet)
ConvNets process images through successive convolutional and pooling layers, then use fully connected layers for classification. The section also introduces transfer learning to address limited target-domain training data.
- CNN overview: CNNs support image classification, object detection, localisation, and segmentation in computer vision.
- Network structure: A neural network maps image inputs to class outputs, with four output nodes representing mould, stain, deterioration, and normal.The described RGB input contains 3×224×224 values.
- CNN layers: ConvNets successively apply convolution and pooling, reducing spatial dimensions while increasing extracted features.Convolution filters generate feature maps; pooling reduces parameters and computation through maximum or average aggregation.
- CNN layers: Fully connected layers connect neurons across successive sub-layers, and the output layer classifies using extracted features.Softmax typically produces a probability from 0 to 1 for each class label.
- Transfer learning: Transfer learning is motivated by deep learning’s need for large labelled datasets, while target-domain data are often substantially smaller than source-domain data.The source and target domains may differ in feature spaces, probability distributions, tasks, or label spaces.
- Transfer learning: Transfer learning reuses knowledge from a source domain and task to improve prediction in a different target domain and task.The paper uses weights from a VGG network trained on ImageNet to classify building-defect images.
4. Methodology
The study develops a transfer-learning CNN model using VGG-16 to classify and localise dampness-related building defects from images. The model uses four image categories and is fine-tuned to adapt pretrained features for defect recognition.
- The model classifies images as mould, stain, paint deterioration, or normal, with paint deterioration covering peeling, blistering, flacking, and crazing.
- The researchers use fine-tuned transfer learning with VGG-16 pretrained on ImageNet rather than training a network from scratch.
- The dataset contains 2622 images collected from mobile phones, handheld cameras, and internet sources, resized into 224×224 thumbnails.
- VGG-16 comprises five convolutional blocks with max-pooling, three fully connected layers, and a final Softmax classifier.
- The original 1 X 1000 classifier is replaced by a 1 X 4 classifier, while early layers remain frozen and only block five is retrained.
- Among examined transfer-learning arrangements, freezing early layers and updating block five performed better than the alternative implementations tested.
Class prediction
The study evaluates classification accuracy and loss for VGG-16, ResNet-50, and Inception models. VGG-16 achieves the strongest reported validation accuracy, while the model also shows class-specific errors and a documented misclassification example.
- 98.86% validation accuracy was achieved by VGG-16, compared with 95.61% for ResNet-50 and 95.42% for Inception.
- VGG-16 reached a final loss of 0.042 on validation, compared with 0.102 for ResNet-50 and 0.144 for Inception.
- The paper documents an example in which the model failed to identify the correct class of damage caused by dampness.
- The model classified mould with an approximately 91% success rate, while stain and paint deterioration achieved approximately 85% and 80%, respectively.
- Overall precision ranged from 82% for deterioration to 89% for stain, while recall ranged from 82% for deterioration to 99% for normal.
Defect localisation using CAM
The model used CAM to localise dampness-related building defects after classifying images, accurately locating many defects while producing some incorrect localisations.
- CAM uses gradients entering the final convolutional layer to produce a coarse map of image regions most significant for class prediction.The method supports localisation in addition to assigning defect classes.
- Representative images showed accurate localisation of paint deterioration, stains, and mould across the displayed examples.The examples are presented in Figure 10.
- Some cases were incorrectly localised, including images containing widespread paint deterioration and a large stain.These examples are presented in Figure 11.
- The section includes a confusion matrix and classification report for evaluating defect-class predictions.
5. Discussion and Conclusions
The study develops a CNN-based system that classifies images into three dampness-related defect categories and a normal category, then applies CAM for localisation. It reports strong classification performance and accurate localisation, while acknowledging limits on defect categories, visibility, and imaging conditions.
- The approach classifies images as mould, stain, paint deterioration, or Normal, with paint deterioration covering peeling, blistering, flacking, and crazing.
- 97.83% training accuracy and 98.86% validation accuracy were recorded after 50 epochs, while the separate 732-image test set achieved 87.50% overall accuracy.The test set contained 183 images per class.
- The test set classified 90% of mould images, 82% of deterioration images, 89% of stain images, and 99% of normal images correctly.
- CAM was integrated to locate defects with high precision, extending the approach beyond classification toward automated detection and localisation.
- Compared with crack-focused studies using dedicated large datasets, the work targets multi-class defects using relatively small datasets.
- The study did not consider multiple defect types simultaneously, invisible defects, or extreme lighting and orientation conditions.The authors identify these as limitations for future work.