The vision system mounted on the extrusion head operates at a high speed of up to 2000 mm/min, necessitating the robust and accurate detection of beads in real-time under varying environmental conditions. To achieve this, selecting an appropriate object detection model based on deep learning is critical. Notably, in this study, the Region of Interest (ROI) was defined as the area of the bead extruded during the AM process. If defined statically, the rapid movement of the extrusion head may cause the actual bead to fall outside the predefined area. Hence, a detection model capable of dynamically identifying ROIs with stability in real-time settings is essential. In this study, several object detection models were compared and analyzed to identify the optimal approach for real-time bead recognition. Object detection involves predicting the class and bounding box coordinates of objects within an image.

To this end, we compared three representative object detection models: Faster R-CNN, SSD-MobileNet V2, and YOLOv8. Object detection models are generally categorized into two-stage and one-stage approaches [32]. Faster R-CNN, a two-stage detector, provides high detection accuracy but suffers from relatively slow inference speed, which restricts real-time applicability [33]. SSD-MobileNet V2, a lightweight one-stage model, achieves fast inference and efficiency, but its accuracy tends to decline under complex backgrounds or low-contrast conditions [34]. YOLOv8, the latest one-stage detector, introduces an anchor-free architecture with improvements in both detection speed and accuracy over YOLOv5 [35,36]. It combines high accuracy with real-time performance, demonstrating robust operation across diverse bead shapes and lighting conditions. The architectural characteristics of these models are summarized in Fig. 5, while the rationale for selecting YOLOv8 as the final model is discussed in the subsequent section.

Although deep learning models generally require large datasets, the images in this study were collected directly from our extrusion-based AM process, ensuring high relevance to the target application. Despite the relatively small scale, a custom dataset of 1400 real bead images was sufficient for model training and validation. The dataset was split randomly into 80% for training and 20% for validation, reflecting a wide range of bead shapes and lighting conditions. Data augmentation techniques such as Random Horizontal Flip and Random Brightness Adjustment were applied to enhance generalization under various lighting and orientation scenarios. While this dataset was adequate for the present material system, if a different polymer, color, or surface finish is used, additional images for fine-tuning would be required to maintain detection accuracy. All models were trained under consistent conditions: Batch Size = 32, Epochs = 100, Optimizer = Adam. During training, both training loss and validation loss were monitored to evaluate overfitting.

In Table 1, mean Average Precision (mAP) represents the detection accuracy of bead regions, FPS (frames per second) indicates the overall processing throughput, and inference time (ms/frame) denotes the computational latency for processing a single image. Together, these metrics provide a comprehensive evaluation of each algorithm’s suitability for real-time monitoring. All models achieved an accuracy of over 0.90, but with notable differences. Faster R-CNN demonstrated the highest accuracy; however, its two-stage architecture resulted in slow inference, making it unsuitable for real-time applications. SSD-MobileNet V2 achieved fast processing speed and was computationally lightweight, making it promising for some real-time scenarios. However, it exhibited unstable performance in low-contrast conditions. YOLOv8 provided a balanced performance in both accuracy and speed, and its modular PyTorch implementation with CUDA optimization enabled seamless integration into real-time monitoring systems.

In addition, the Receiver Operating Characteristic (ROC) curves shown in Fig. 6 further validate these findings. Faster R-CNN achieved the highest area under the curve (AUC ≈ 0.947), indicating superior detection accuracy, but its low FPS limits real-time applicability. SSD-MobileNet V2 yielded a slightly lower AUC (≈0.914), yet maintained the fastest inference speed. YOLOv8 achieved an AUC of approximately 0.942, demonstrating both high accuracy and real-time speed, thereby confirming its role as the most balanced and practical model for dynamic extrusion-based additive manufacturing.

Based on the results, YOLOv8 was selected as the final object detection model for this study. It provided the best trade-off between bead recognition accuracy and real-time processing speed. Furthermore, its modular architecture, GPU acceleration support, and adaptability to various lighting and shape conditions, enabled by data augmentation, demonstrated superior robustness and applicability in real-world manufacturing environments. YOLOv8 also showed fast convergence and stable training curves without overfitting, validating its suitability for deployment in dynamic extrusion-based additive manufacturing settings.

In this study, an image pre-processing and analysis algorithm was designed to dynamically update the Region of Interest (ROI) based on bead recognition and precisely measure bead width within the extracted ROI. The ROI was defined to cover only the local deposition area around the bead (approximately 10 mm × 30 mm in physical scale, based on camera calibration), ensuring that the contour was fully included while excluding unnecessary background. A key objective of the algorithm is the accurate detection of bead contours, for which a lightweight implementation based on the Canny Edge Detection algorithm was developed.

Canny Edge Detection is a traditional edge extraction method that offers both robustness and computational efficiency [37,38]. It ensures stable contour detection even under complex backgrounds and, due to its fast processing speed, is highly suitable for real-time additive manufacturing environments. As illustrated in Fig. 7a, the Canny Edge algorithm was applied to isolate bead contours effectively.

To further enhance contour accuracy, Bilateral Filtering was applied prior to edge detection [39]. Unlike conventional Gaussian smoothing, Bilateral Filtering removes background noise while preserving sharp edges, which is advantageous for extracting bead boundaries without blurring. This effect is clearly shown in Fig. 7b, where Bilateral Filtering provides superior edge preservation compared to Gaussian Blur, thereby improving the reliability of bead contour extraction.

The complete bead width measurement process is illustrated in Fig. 8. The algorithm first applies Bilateral Filtering to the input image within the dynamically detected ROI. Next, Canny Edge Detection is used to extract the bead contours. The pixel distance between detected contour edges is then calculated and multiplied by a calibration coefficient (pixel-to-mm ratio) to derive the final bead width. This coefficient was obtained through a camera calibration process using a calibration board, where physical measurements were compared with corresponding pixel distances.

The entire algorithm was implemented in Python with the OpenCV library, and the average processing time was measured at approximately 2.2 ms per frame, satisfying real-time video processing requirements. In conclusion, the proposed algorithm provides a reliable and efficient approach for bead contour extraction and width estimation in dynamic AM environments. It simultaneously ensures edge detection accuracy, real-time responsiveness, and computational efficiency, while serving as a foundation for future extensions such as separating outer and inner contours for layer-by-layer quality analysis in large-scale additive manufacturing.

In this study, inspired by previous research indicating that nozzle feed speed and material extrusion speed significantly influence bead geometry characteristics, such as width, thickness, and surface quality, these two variables were selected as the primary process parameters [40]. The objective was to quantitatively analyze changes in bead geometry resulting from variations in the ratio of these parameters. In particular, it was observed that excessively high feed speeds or insufficient extrusion speeds result in under-extrusion, which can lead to bead breakage or reduced thickness. Conversely, excessively high extrusion speeds result in over-extrusion, leading to excessive bead formation or degradation in surface quality. Based on these observations, the process conditions were systematically designed with a focus on the interaction between feed speed and extrusion speed, and the sensitivity of bead geometry to these parameters was quantitatively assessed through experiments. Prior to the experimental design, considerations were given to minimizing material consumption and experimental costs by reducing the number of trials while still accurately quantifying changes in bead geometry and identifying the optimal process conditions. Consequently, Central Composite Design (CCD) was adopted as the experimental design method for process optimization. CCD, a technique within the response surface methodology (RSM), allows for the analysis of linear, nonlinear interactions, and quadratic effects with relatively few experiments. In this study, 13 experimental conditions were established for the two variables, feed speed and extrusion speed, including center points, axial points, and factorial points. The selected experimental conditions are shown in Table 2. The levels of each variable were set according to the fundamental principles of CCD at minimum (−α), center (0), and maximum (+α) values, with the α value set at 1.414 to ensure repeatability and normality of the experiments.

This design approach offers superior predictive power with fewer experimental runs compared to conventional full factorial designs, making it suitable for analyzing complex and highly nonlinear AM process variables. The experiments employed ABS + chopped CFRP (20 wt%) composites, the same material as the planned target part. The material underwent a pre-print drying process to minimize moisture content, and the deposition temperature was fixed at 260°C. Other parameters, including nozzle diameter, layer height, bed temperature, and ambient temperature, were deliberately kept constant (see Table 3) to ensure experimental reliability and eliminate their influence as variables. This was consistent with the study’s focus on verifying the applicability of the proposed real-time monitoring algorithm, for which feed speed and extrusion speed were selected as the primary variables affecting bead geometry.

Furthermore, as illustrated in Fig. 9, a test pattern that accentuates bead geometry was designed and printed for each experimental condition to facilitate clear visual observation. Beads produced under each process condition were independently deposited, and a vision-based monitoring system was developed to automatically measure bead width.

To validate the accuracy and reliability of the developed vision-based bead width measurement monitoring system, a comparative analysis was conducted between bead width data collected from the system and physical measurements. Experiments were conducted under 13 different printing conditions, and bead widths of the fabricated specimens were directly measured using a Vernier caliper, which offers a high measurement precision with an error margin of ±0.02 mm. The mechanically obtained measurements served as reference values for the vision-based system. To ensure data consistency, measurements were repeated at the same location at least three times for each experimental condition, and the average values were calculated and compared with the measurements obtained from the vision system. The mean relative error was 2.36% (max. 7.1%), with most conditions within 5% (Table 4). This level of accuracy implies that the system can detect bead width deviations with high reliability, thereby enhancing defect detection and increasing confidence in process control.

Guided by volume conservation, the bead cross-section A transported at the travel speed vf must balance the deposited flow Q with efficiency η(Avf=ηQ) [41]. With A≈κwh (shape factor κ, layer height h) and Q≈γRPM (screw speed as extrusion setting), the scaling becomes: (3)w=KrpmRPMvf,Krpm=ηγκh

Eq. (3) quantifies how bead width w increases with the flow-to-speed ratio; RPM is the screw speed (rev/min) and vf is the feed (mm/min). Using the vision-derived widths in Table 4 as the response and the CCD settings (13 runs) for the predictors, we estimate Krpm by origin-constrained least squares: (4)zi=RPMvf,K^rpm=∑iziwi∑izi2

This yields K^rpm=9.19×102mm2/rev. Model adequacy on the 13 runs is Spearman’s ρ = 0.734 (p = 0.0043) and LOOCV RMSE ≈ 3.80 mm (≈19% of the mean width), indicating trend-level accuracy adequate for deriving operating set-points rather than fine prediction. (5)(RPMvf)∗=w∗K^rpm

Accordingly, Eq. (5) provides a compact relation to compute the operating ratio for a target width, while substituting K^rpm into Eq. (3) gives a coarse width estimate for given RPM and vf within CCD design space. As real-time vision data continue to accumulate, Krpm can be updated online and extended with additional covariates to further improve accuracy without sacrificing physical interpretability.

Despite being based on a camera-driven vision measurement approach, the system consistently maintained stable contour recognition and bead width measurement performance, exhibiting minimal susceptibility to environmental variables such as lighting fluctuations, post-extrusion thermal deformation, and residual material accumulation around the nozzle. Validation using the CCD-based experimental configuration confirmed that the vision-derived bead width measurements closely matched physically measured values, demonstrating the high reliability of the system’s quality assessment capabilities. In contrast to conventional bead quality evaluation methods, which often require halting the process or destructively sectioning completed specimens for physical measurement, thereby disrupting production flow, complicating repeated assessments, and introducing variability due to measurement timing, the developed vision-based real-time monitoring system enables continuous, non-intrusive data acquisition and analysis without interrupting the additive manufacturing process. This capability not only facilitates immediate quality feedback and process control for bead geometry but also offers significant improvements over traditional manual measurement approaches in terms of precision, productivity, and operational practicality. As such, the proposed system represents a rapid, accurate, and fully automated alternative for measuring bead width, with strong potential to enhance real-time quality assurance in additive manufacturing applications.

To verify the applicability of the developed vision-based bead monitoring system in real-world environments, additive manufacturing experiments were conducted using an actual aerospace component. The test subject was the rear cockpit frame structure of the KF-21 fighter jet. Stable deposition process conditions, derived from prior simulation data and experimental analysis, were applied to fabricate the part using 3D printing. Notably, the component included curved and complex outer geometries, making the uniformity of extrusion a critical factor. During the experiment, the vision-based monitoring system was deployed in real-time under the predefined process parameters. The system was configured to trigger alerts upon detecting bead geometry anomalies, such as width imbalances, discontinuities, or deviations, allowing the operator to identify and respond to issues promptly. This implementation not only ensured real-time process stability but also established a continuous quality monitoring system without interruptions, demonstrating the system’s effectiveness in a real manufacturing environment.

The material used for the experiments was the same ABS + chopped CFRP (20 wt%, Samyang Corporation) composite applied in earlier tests. The deposition speed was set between 70 and 90 mm/s to reflect actual industrial processes, considerably faster than typical FDM methods. Such high-speed deposition often leads to increased geometric instability and bonding defects due to insufficient cooling of the material before subsequent bead deposition. To address these challenges, the deposition strategy was modified compared to that used in previous trials. Specifically, the number of beads increased from one to four, and the CAD model wall thickness was adjusted to fill potential void regions.

Repeated experiments under varying bead width conditions revealed that a bead width of 12 mm provided the most uniform deposition geometry with the lowest porosity. This conclusion was based on real-time contour data collected by the vision system. Through repeated trials, this condition exhibited the highest consistency in shape replication and the least variability in porosity ratio, offering a quantitative foundation for future process optimization and feedback control design.

The results of applying the proposed vision-based bead width monitoring system to the deposition process of the KF-21 rear cockpit component are shown in Fig. 10. Fig. 10a displays the real-time image data of the beads, illustrating how the model automatically detects the region of interest (ROI) and measures bead width as the extrusion head moves. The system achieved real-time analysis at 30 FPS, with bead widths recorded frame by frame to allow tracking of shape variations over time.

Fig. 10b shows that the measured bead widths remained close to the target value of 12 mm throughout the frames, though minor anomalies were detected in some frames. For instance, temporary reductions in bead width were observed due to momentary decreases in extrusion speed or slight bead oscillations caused by head vibrations. These irregularities were identified through deviation patterns in quantitative width data and could be integrated into a real-time alert system by setting tolerance thresholds.

Fig. 10c presents the complete KF-21 printed component, demonstrating the accuracy of the shape and the uniformity of bead patterns. To obtain precise dimensional measurements of the part, a 3D shape scan was performed using an Absolute Arm laser scanner (8525-7/RSS, Hexagon, Stockholm, Sweden), a high-precision articulated measuring device with a repeatable precision of ±0.03 mm, commonly used for industrial dimensional inspections. Fig. 10d compares the scanned geometry with the CAD model, revealing that the overall geometric deviation remained within 6 mm. This indicates that despite the effects of shrinkage and material flow characteristics inherent to large-scale composite additive manufacturing, the dimensional accuracy remained within practical limits.

Through the additive manufacturing experiments using a complex-shaped aerospace component, the applicability and technical validity of the vision-based bead monitoring system were empirically demonstrated. Even under modified deposition conditions involving four beads, the proposed system successfully recognized each bead individually. It precisely measured their widths in real time, highlighting its scalability and adaptability to multi-bead structures. Furthermore, the algorithmic architecture and sensor parameters established during preliminary experiments proved effective in the actual deposition process, demonstrating the system’s potential as a core infrastructure for real-time quality control and feedback-based process management.