Inspection of Concrete Structures by a Computer Vision Technique and an Unmanned Aerial Vehicle

We have proposed a visual inspection technique for concrete structures using deep learning and a hardware ecosystem, an Unmanned Aerial Vehicle (UAV). The UAV is a quadcopter that can fly to unreachable sections of a site which consists of a camera that captures images of the concrete surfaces via a mobile device and feed the real time images in the CNN model. The images taken from such remote locations may contain different types of surfaces, shadowed regions and surfaces with holes. The cracks are properly detected by the CNN `AlexNet' algorithm and masking with sliding window technique in such conditions due to variation in the image data set. The experimental results were simulated on a standard online data set of 40,000 images of Mendeley Data which is freely available and 3000 images have been chosen from the entire data set for this method. The classes have been divided into 2 categories of `crack' and `no crack' for the proposed method's data set. There are 1050 training images and 450 testing images for each category. Experimental results were achieved on Google Colab cloud service using Python Tensorflow API (Application Programming Interface). The proposed `AlexNet' CNN algorithm achieves 98.4 % accuracy and the model is deployed to a masking technique with sliding window to detect cracks in a 3008×2000 pixel resolution image by breaking the image into 227×227 pixel resolution image patches. The experimental results have proved that the proposed method handles noisy background such as cracks with shadows and stains, cracks on rusty and rough surfaces and minor dimension cracks with good efficiency.