Abstract:

Objective Addressing the issues of low efficiency, high subjectivity and unstable accuracy of traditional size measurement methods in the apparel industry, as well as the high cost and complex operation of existing three-dimensional measurement devices that hinder widespread adoption, this paper proposes an automatic measurement method for women's trousers that integrates machine vision with the YOLO11 deep learning model. The approach aims to achieve non-contact precision measurement of key sizes of women's trousers, thereby advancing digital and intelligent transformation in garment size inspection, while providing a robust technological solution for automated quality control in the apparel industry.

Method The measurement system comprised hardware and software components, where the hardware employed devices for image acquisition and the software performed size measurement. A dataset comprising 2000 images of women's trousers was constructed for the experiment. After image preprocessing, image annotation and format conversion, YOLO11n-pose pose estimation model was used to detect key points. The pixel coordinates of the detected key points were converted from the pixel distance to the actual physical distance using Zhang's camera calibration method, enabling calculation of sizes for five target regions. Three different women's trousers were selected to compare the automatic measurement results with the manual measurement results, and the accuracy of the measurement was evaluated by absolute error and relative error.

Results The experiment used the YOLO11n-pose key point detection model to train on a self-built dataset of women's trousers. Experimental results demonstrated a gradual decrease and eventual stabilization of the overall loss value, indicating good fitting performance. As training epochs increased, the model's precision, recall, and mAP(mean Average Precision) values gradually rose and became stabilized, indicating improved model performance with convergence. Specifically, the model achieved 100% precision and recall, with mAP50 and mAP50-95 reaching precision of 99.5% and 99.3% respectively. From the detection results, it was seen that the model acquired the ability to effectively detect key measurement points of women's trousers through learning a large amount of labeled data, providing reliable support for subsequent size measurement. In the camera calibration section, calibration plates were created and images of the calibration plates were captured to determine the pixel equivalent. The average reprojection error after calibration was 0.070 549 pixels, indicating high-precision calibration that meets experimental requirements. The pixel equivalent was ultimately calculated to be 0.298 913 mm/pixel. After converting pixel distances to actual physical distances, the sizes of women's trousers were calculated based on the vertical and horizontal Euclidean distances between two key positioning points. Three pairs of women's trousers were selected for comparing the automatic measurements with the manual measurements. The evaluation criteria selected were the absolute error and relative error. The average absolute errors between the measured and the actual sizes of the five measurement locations on the three pairs of women's trousers were calculated to be 5.45 mm, 5.89 mm, 1.27 mm, 2.35 mm, and 6.26 mm, respectively. The average relative errors were 0.82%, 0.61%, 0.58%, 0.64%, and 0.64%, respectively. The measurement results were found to comply with the industry standard requirements, indicating that the automatic measurement method for women's trouser sizes proposed in this experiment, based on machine vision and deep learning, has high accuracy and reliability.

Conclusion This study combines machine vision, automatic measurement system design, and the lightweight deep learning algorithm YOLO11n-pose to achieve non-contact precise measurement of key sizes of women's trousers. Experimental results show that the system's measurement error for women's trousers is consistently within the industry's acceptable range, validating the feasibility of this method in actual production. From an application perspective, this system breaks through the bottleneck of traditional manual measurement, providing apparel companies with an automated and standardized size inspection solution. In the future, the measurement robustness can be further improved, and the deep application of machine vision technology in size inspection across all apparel categories can be explored to drive the digital transformation of size measurement processes within the apparel industry.

Key words: woman's trouser, machine vision, deep learning, camera calibration, YOLO11, key point detection, trouser size, non-contact measurement