Predicting the defects in wooden structures by using pre-trained models of Convolutional Neural Network and Image Processing

摘要: Wood is the oldest material used in construction industry and wooden structures are often exposed to harsh conditions of environment. This exposure lead to deterioration due to mechanical and weathering effect. In this study, three different pre-trained model of Convolutional Neural Network (CNN) are trained for predicting the defects in wooden structures and comparative study of CNN models. The dataset of 5000 wooden images are collected from online available sources and site visits. The 80% images of dataset are defected and remaining 20% are un-defected. The defected dataset is further equally divided into four categories Horizontal Crack, Vertical Crack, Diagonal Crack, and Knots in wooden structures. The pretrained models of CNN ResNet18, ResNet50, and ResNet101 are trained and validate on this dataset to check the accuracy, precision, Recall, and F1 Score of predictions and effect of deep layers on them. The Computational complexity of each CNN models also assessed during training, Validation, and testing. The Confusion Matrixes and overall performance of CNN models are also generated for comparative study and found ResNet50, the best model for the prediction of defects in wooden structures..