Abstract:

Objective Conventional manual stapling methods for measuring cashmere fiber length suffer from low efficiency and subjectivity, requiring 30-40 min for each sample. Existing automated instruments, including Almeter capacitors and optical fiber diameter analyzers (OFDA), have limitations such as incomplete short fiber clamping and restriction to specific sample forms. With computer vision-based methods on the other hand, the threshold segmentation frequently causes fiber breaks, and tracking individual fibers in complex intersecting networks is extremely difficult. This study develops an automated measurement method based on an improved U-Net and global optimization algorithms to overcome these limitations.

Method An improved U-Net (encoder-decoder convolutional network) architecture incorporating directional convolution, spatial attention, nonlocal feature modules, and morphological processing layers was constructed to repair fiber breaks caused by threshold segmentation. A UNet++ (nested U-Net) model with an EfficientNet-B0 encoder and a hybrid Dice-focal loss function was built for crossover region detection. A two-stage globally optimized path tracking algorithm using directional similarity, historical experience scores, and necessity coefficients, combined with a simulated annealing-based matching algorithm, was adopted to obtain the complete fiber length distribution.

Results The improved U-Net achieved a recall of 98.67% and an F1 value of 91.29% in fiber break repair. Precision was reduced due to morphological dilation operations required to ensure full fiber connectivity, as these operations cause slight over-prediction at the pixel level. However, this trade-off ensures that fiber breaks are fully repaired, reducing tracking failures in subsequent steps. The crossover detection model based on UNet++ achieved an intersection over union (IoU value) of 71.74% and a recall of 85.42% on the test set, effectively identifying overlapping regions where pixel loss occurs.

Measurements were conducted on cashmere samples from five batches provided by an enterprise in Inner Mongolia and compared with manual stapling results. The absolute error ranged from 0.23 to 1.07 mm across all sample groups, with an average absolute error of 0.67 mm and an average relative error of 1.78%. Three of the five sample groups had relative errors below 2%, and the remaining two were within 3%. One sample group showed a larger deviation, attributable to an insufficient number of medium-length fibers in the sample, which limited the convergence of the simulated annealing-based matching process.

The two-stage path tracking algorithm was validated in complex crossover scenarios. In a case involving 12 crossovers, the algorithm correctly identified the exit direction by incorporating historical decision weighting, even when directional similarity scores of competing directions differed by only 0.17. In a high-complexity case involving 30 crossovers, where the most challenging decision point had 6 candidate exit directions, the algorithm selected the direction consistent with the actual fiber extension by combining directional similarity, historical experience scores, and global optimization weighting.

The overall fiber length range and distribution pattern were consistent with those of the manual stapling method. The frequency of fibers below 10 mm was generally higher than in the manual method, consistent with the known limitation that short fibers in manual stapling are densely packed and difficult to separate accurately.

Conclusion An automated cashmere fiber length measurement system integrating deep learning and intelligent optimization algorithms was developed. The improved U-Net architecture incorporating directional convolution, spatial attention, and nonlocal feature modules achieved a F1 value of 91.29% and an IoU value of 91.29% of 80.70% in fiber break repair. The two-stage globally optimized path tracking algorithm enabled accurate single-fiber tracking in complex crossover networks by incorporating directional continuity, historical decision experience, and global coordination among competing fiber paths. The simulated annealing-based matching algorithm obtained the complete fiber length distribution. Evaluated on five batches of cashmere samples, the system showed good agreement with the manual stapling method, with an average absolute error of 0.67 mm and an average relative error of 1.78%. The method addresses the technical challenges of threshold-induced fiber breakage and individual fiber tracking in complex networks, and may be extended to other natural fiber types in future work.

Key words: cashmere, fiber length measurement, deep learning, image processing, tracking algorithm