Investigation of steel frame damage based on computer vision and deep learning

Visual damage inspection of steel frames by eyes alone is time-consuming and cumbersome; therefore, it produces inconsistent results. Existing computer vision-based methods for inspecting civil structures using deep learning algorithms have not reached full maturity in exactly locating the damage. This paper presents a deep convolutional neural network-based damage locating (DCNN-DL) method that classifies the steel frame images provided as inputs as damaged and undamaged. DenseNet, a DCNN architecture, was trained to classify the damage. The DenseNet output was upscaled and superimposed on the original image to locate the damaged part of the steel frame. The DCNN-DL method was validated using 144 training and 114 validation sets of steel frame images. DenseNet, with an accuracy of 99.3%, outperformed MobileNet and ResNet with accuracies of 96.2% and 95.4%, respectively. This case study confirms that the DCNN-DL method effectively facilitates the real-time inspection and location of steel frame damage. • Deep learning approaches is proposed to locate the damaged part of the steel frame. • DenseNet based deep convolutional neural network is implemented and validated. • Data augmentation and Grad-CAM visualization techniques is implemented. • The proposed model accurately locates the damages in the steel structures.