Rheological properties of cement paste are critical factors governing the workability of fresh concrete, significantly affecting construction performance, structural safety, and durability. Although rheometers provide direct measurements of parameters such as yield stress and plastic viscosity, their field application is often restricted by equipment costs and operational complexity. Consequently, the mini-slump flow test remains a prevalent method for evaluating workability due to its simplicity. Theoretical frameworks allow for yield stress calculation based on the final spread diameter. However, conventional analysis relies solely on the static final value, neglecting the dynamic flow process. This limitation results in a failure to capture dynamic information related to plastic viscosity, which dictates the spreading rate and flow duration. Recent computer vision technology has demonstrated a broad potential in construction materials research. Some previous studies attempted to record flow spread using cameras, they often suffered from reliance on manual data processing, insufficient precision, or requirements for expensive high-speed imaging equipment. To address these limitations, this paper was to propose a high-precision characterization method for the flow spread process of cement paste based on computer vision. This study could achieve automated, dynamic, and quantitative analysis of paste flow via integrating the You Only Look Once version 11 (YOLOv11) object detection model with the Segment Anything Model 2 (SAM2), and employing a dual geometric correction system, providing a robust data foundation for inverting rheological performance.

P·O 42.5 ordinary Portland cement, fly ash, and slag were utilized to prepare 80 groups of cement pastes with varying water-to-binder ratios (i.e., 0.45, 0.50, and 0.55) and mineral admixture dosages ranging from 0% to 50%. A polycarboxylate superplasticizer was added to regulate fluidity, creating a dataset with diverse rheological properties. The image acquisition setup consisted of a smartphone camera fixed 400 mm vertically above a 10 mm thick tempered glass plate, with an A3 paper sheet placed underneath as a spatial reference, recording videos at 1920 × 1080 resolution. The proposed characterization method followed a sequential pipeline comprising four stages, i.e., perspective correction, automated segmentation, refraction correction, and dynamic parameter calculation. First, to address trapezoidal distortion caused by oblique camera angles, a global homography matrix was calculated via detecting the four corners of the A3 reference paper in the final video frame. This matrix was applied to the entire video sequence to restore an orthogonal top-down view. Subsequently, a combined deep learning approach was employed for segmentation, where YOLOv11 detected the mold position in the initial frame to generate a "point prompt" that initializes SAM2. Leveraging its Video Object Segmentation (VOS) capability and memory mechanism, SAM2 automatically tracks and segments the paste mask throughout the sequence, effectively resolving occlusion issues where the mold or operator hands could block the paste during the initial lifting phase. Following segmentation, a refraction correction model based on Snell's law was applied to eliminate the visual displacement caused by light passing through the glass plate, establishing an accurate physical relationship between pixel width and true scale. Finally, the spread diameter D(t) and instantaneous velocity V(t) curves were extracted to quantify dynamic behavior based on the corrected masks.

The experimental validation confirms the high efficacy of the proposed method in both automated segmentation and measurement accuracy. In terms of segmentation, the prompt-based mode utilizing YOLOv11 and SAM2 proves superior to the automatic mask generation mode. The model effectively propagates mask information from clear final frames back to initial frames via leveraging the VOS function, generating continuous and precise contours even when the paste is partially occluded. Regarding measurement precision, the constructed dual correction system significantly mitigates errors. The experimental results indicate thatan optical distortion causes a maximum relative error of 4.1% without a refraction correction. The application of the refractive correction model reduces this error to within 1.6%. A validation study involving 80 groups of pastes shows a linear correlation between the calculated final diameters and manual measurements, with a coefficient of determination (R²) of 0.999, a Mean Absolute Error (MAE) of 0.87 mm, and a Mean Absolute Percentage Error (MAPE) of merely 0.49%. Based on the accurate extraction of D(t) and V(t) curves, four dynamic parameters were defined, i.e., final spread diameter (D_max), peak velocity (V_max), flow stabilization time (ts_table), and time to reach peak velocity (t_Vmax). Crucially, the analysis demonstrates that distinct pastes can exhibit nearly identical final diameters yet possess significantly different peak velocities and flow time. This discrepancy indicates that static measurements alone fail to capture variations in plastic viscosity and flow dynamics. The proposed dynamic indicators effectively capture these distinctions that conventional static tests fail to distinguish, offering a more comprehensive description of flow behavior.

This study established an automated segmentation method for cement paste via integrating YOLOv11 and SAM2, which could effectively solve the dependency on external prompts and address occlusion problems during the initial flow stage through a memory-based mask tracking mechanism. Furthermore, a dual error correction system incorporating perspective and refraction corrections was constructed to quantitatively rectify geometric and optical distortions, reducing the maximum relative measurement error from 4.1% to less than 1.6% and ensuring high fidelity of physical dimensions. Based on these foundations, a dynamic quantitative analysis method was proposed by defining key parameters such as D_max, V_max, t_stable, and t_Vmax based on the dynamic curves. Compared with conventional methods, this dynamic characterization could offer improved information density and automation, achieving a more comprehensive characterization of cement paste flow behavior through quantitative analysis of the flow process. Ultimately, the method could provide a technical support for the automated and intelligent testing of cement paste workability.