Concrete crack segmentation using U-Net based models with various pre-trained backbones

Abstract

The use of deep learning to detect and analyze structural damage in concrete is gaining increasing attention, but exploration of U-Net with various pre-trained backbone models has not been conducted. This study aims to overcome these limitations by developing a U-Net-based concrete crack segmentation framework that utilizes eight modern and lightweight backbones, namely EfficientNetB0, VGG16, VGG19, ResNet34, ResNet50, DenseNet121, MobileNetV2, InceptionV3, and Xception. A total of 458 labeled images from the Concrete Crack Segmentation dataset were used as training and testing data, representing the variety of concrete surface conditions in real environments. Evaluation was conducted using Dice and Intersection over Union (IoU) metrics. The test results showed that the Xception backbone provided the best performance, with a Dice value of 0.8461 and an IoU of 0.7384, surpassing EfficientNetB0 and VGG16/VGG19 which were previously considered stable in segmentation tasks. This finding confirms that the depthwise separable convolution mechanism in Xception is capable of extracting thin crack features more representatively. This study provides an important contribution in selecting the optimal backbone for concrete crack segmentation models, while opening up opportunities for implementing more accurate and efficient structural condition monitoring technology on an industrial and public infrastructure scale.