Intelligent quality assessment of concrete vibration using computer vision and large language models

Abstract

The monitoring of concrete vibration quality is crucial for ensuring construction quality. This paper proposes a monitoring method that combines computer vision and Large Language Model (LLM). First, an unsupervised shadow removal method is used to optimize image quality. Next, a multi-head classification model is applied to conduct a multi-dimensional comprehensive assessment of vibration quality. After that, the classification results are mapped to natural language information through a key-value image-to-text mapping method. Finally, the natural language is used for inference in the LLM to generate real-time feedback. Experimental results show that the proposed method achieves an accuracy of 94.45 % in classifying the vibration quality. Additionally, by combining image classification results with LLM for logical reasoning and feedback generation, the system can provide detailed descriptions of compaction quality and corresponding solutions. This research has been successfully applied in real-world projects and is expected to promote the intelligent development of construction operations.