A hybrid deep learning method for the crack detection and classification in surface images

ABSTRACT Structural Health Monitoring (SHM) is critical to the safety, sustainability, and reliability of the infrastructural system because it allows regular evaluation of the integrity of the structure. Surface cracks are one of the main signs of degradation in structural elements that can affect load-bearing capacity, causing serious damage in the case of undetected problems in early stages. In order to overcome problems inherent in manual inspections and traditional image processing methodologies, the proposed work develops a deep learning (DL) model for automatic detection of surface cracks. This methodology uses Gaussian filter as a preprocessing step, snake as a segmentation method, EfficientNet as feature extractor, and Bidirectional Long Short-Term Memory (Bi-LSTM) as classifier of cracked and uncracked surfaces. The proposed algorithm was trained on a dataset comprising images related to cracks and non-cracks from the Kaggle database. The performance of the model was measured using commonly used evaluation criteria, which include accuracy, precision, and recall. It was found through experimentation in Python that the efficiency of the proposed model based on EfficientNet-BiLSTM architecture is 92.10% in terms of accuracy, along with good performance in precision and recall, surpassing the conventional transfer learning models, such as VGG16, Inception, and DenseNet.