【摘要】 针对传统羊绒纤维长度测量方法效率低、精度受人为因素影响的问题，采用卷积神经网络技术，对羊绒纤维在阈值化处理中的断裂修复和交叉区域检测进行系统研究；设计了基于U-Net框架的纤维补断模型和基于ResNet-50编码器的交叉点检测模型，通过引入跳跃连接机制和混合损失函数，解决因阈值处理导致的纤维断裂和类别不平衡的问题。研究表明，纤维补断模型的像素准确率达到98.72%,交叉点检测模型具有良好的分割精度和泛化能力；模型测得的纤维平均长度与手排法结果的标准差最低为0.28 mm,验证了该方法在羊绒纤维长度测量中具有较高的精度、稳定性和实用性。更多还原

【Abstract】 To address the issues of low efficiency and human factor-induced inaccuracies in traditional cashmere fiber length measurement methods, convolutional neural network technology was adopted to conduct a systematic study on fracture repair and cross-region detection during the thresholding process of cashmere fibers. A fiber fracture repair model based on the U-Net framework and a cross-point detection model based on the ResNet-50 encoder were designed. By incorporating skip connection mechanisms and hybrid loss functions, the problems of fiber continuity restoration and class imbalance were effectively resolved. The research demonstrates that the fiber fracture repair model achieved a pixel accuracy of 98.72%, and the cross-point detection model exhibited strong segmentation precision and generalization capability. The standard deviation between the average fiber length measured by the model and the results from the manual arrangement method was as low as 0.28 mm, verifying that this method offers high precision, stability, and practicality in cashmere fiber length measurement.更多还原