Vision method based on deep learning for detecting concrete vibration quality

Abstract

The vibration quality of concrete is crucial to ensure the long-term safe operation of concrete structural components. In recent years, computer vision technology based on deep learning has achieved excellent results in the field of concrete quality inspection. However, when used for assisting construction inspection, this technology relies heavily on large-scale, labeled, high-quality image data. To solve this drawback, this study proposes a vision-based method that integrates semi-supervised learning and data augmentation for detecting the concrete vibration quality. Initially, StyleGAN2 was adopted as the data augmentation strategy to improve the diversity of the dataset. Then, SE-ResNet50, a model that couples an attention mechanism module and residual network, was employed as a classifier for accurately extracting information contained in images. Subsequently, in order to reduce the workload of data annotation, a novel semi-supervised learning method (Co-MixMatch) was proposed to train the model by coupling MixMatch with co-training. Finally, the trained model was deployed on mobile devices to assist onsite construction workers in detecting the quality of concrete vibration. A real-world concrete dam dataset was employed to verify the proposed method. Based on the experimental results, the proposed method improves the accuracy of the baseline method by 3.62% on average. Additionally, the proposed method achieves an accuracy of 0.9600, which is only 0.67% lower than that of supervised learning, while only requiring 20% of the labeled data. Therefore, the proposed method has great application prospects and can further promote the intelligent development of concrete vibration inspection.