CRACK DETECTION ON CONCRETE SURFACE BY DEEP LEARNING FROM VGG16 ARCHITECTURE

Abstract

            Crack is one of the main problems in concrete structure, which is caused by weathering, rust in reinforcing steel, collapse, erosion of structures. Structural inspection is important to verify the usability of structure. Automated crack detection is an essential tool to help improving efficiency in inspection systems. This research presents the methods for automatic crack detection system using a pre-trained convolutional neural network model by transfer learning technique. The transfer learning technique can be customized a pre-trained model with (i) feature extraction, and (ii) fine-tuning. The VGG16 is one of the Convolutional neural network architecture, which is a pre-trained model that will be used to detect cracks on concrete beam images in this paper, and compared results with Spatially Tuned-Robust Multi feature (STRUM features) with the AdaBoost classifier.

            The classification performance of VGG16 by fine-tuning method is 95.76% accurate and the VGG16 by feature extraction method is 84.56% accurate. The STRUM classification with the AdaBoost classifier is 71.85% accurate.