Airtightness, the ability of a product to prevent gas leakage, is a critical performance indicator for high-performance equipment in industries such as aerospace, shipping, and chemical engineering. During the manufacturing process, airtightness testing of key components is essential. The liquid application method is widely used in batch manufacturing due to its low cost and operational simplicity. Traditional manual inspection, although a common method, is highly subjective, difficult to quantify, and inefficient. It also carries a high risk of missed or incorrect judgments due to inspector fatigue. To address these limitations, this study proposes a quantitative airtightness testing system based on the liquid application method and computer vision technology to transition from manual observation to objective, automated detection.

The proposed system, developed using the PyQt5 framework, integrates image acquisition, parameter configuration, camera calibration, intelligent soap-bubble detection, bubble size measurement, and result visualization. It runs on a platform equipped with a GTX1050Ti graphics card and supports up to six cameras simultaneously for comprehensive multi-view surface coverage. For accurate physical measurement, a camera calibration method using standard spheres is employed. By placing 5 mm standard spheres on the product surface, the system applies Canny edge detection and Hough circle transform to extract sphere contours and compute local pixel-to-millimeter ratios. To compensate for surface curvature and lens distortion, a cubic spline interpolation algorithm is used to establish a global mapping across the detection area. For bubble detection and segmentation, particularly for bubbles with irregular shapes, varying sizes, and complex backgrounds, a lightweight YOLOv11 instance segmentation network is adopted. A dataset of more than 1,200 images collected from real experimental environments is constructed for training. Through transfer learning and iterative optimization, model parameters are reduced to approximately 40% of the standard version while maintaining high detection accuracy. Following segmentation, the maximum chord length method calculates the equivalent diameter of each bubble, avoiding inaccuracies associated with traditional equivalent circle assumptions.

Experimental validation was conducted on real product components with known airtightness defects. A total of 627 images were collected from six viewing angles. The system demonstrates stable detection and tracking of soap-bubble generation and evolution. It achieves a detection accuracy of 95.3% for bubbles exceeding a predefined threshold, with average and maximum measurement errors of 0.19 mm and 0.30 mm, respectively. The optimized YOLOv11 model maintained high segmentation accuracy while achieving an inference speed of over 15 frames/s. The multi-camera configuration and calibration method ensure reliable measurement consistency across the detection area, effectively supporting quantitative leakage analysis.

The proposed system transforms the traditional subjective soap-bubble inspection method into an objective, quantitative, and automated approach for airtightness testing. By integrating multi-camera imaging, precise calibration, deep learning-based bubble segmentation, and the maximum chord length method, the system substantially improves detection reliability.