郭瑞桁1 郭瑞洲1 刘夏丽2
（1.广东瑞洲科技有限公司，广东 佛山 528200
2.广东工业大学，广东 广州 510006）
摘要：针对皮革表面缺陷区域定位及柔性裁片形变引起的CAD路径偏移问题，提出一种改进SAM的皮革缺陷分割与自适应裁切方法。该方法基于SAM-Med2D ViT-B图像编码器、通道-空间双分支适配器构建Leather-SAM，结合ORB特征匹配与全局仿射变换矩阵、多横梁裁切安全控制等，构建基于Leather-SAM的缺陷识别与裁切优化算法。实验结果表明，该方法的IoU、Dice和轮廓F1分别为54.58%、65.22%和88.08%，推理速度为12.2 f/s，为皮革智能排版与裁切提供了可复现的技术流程。
关键词：皮革缺陷；SAM-Med2D；二值分割；仿射校正；自适应裁切
中图分类号：TP391.41 文献标志码：A 文章编号：1674-2605(2026)04-0008-08
DOI：10.12475/aie.20260408 开放获取
An Improved SAM for Leather Defect Segmentation and Adaptive Cropping
GUO Ruiheng1 GUO Ruizhou1 LIU Xiali2
(1.Guangdong Ruizhou Technology Co., Ltd., Foshan 528200, China
2.Guangdong University of Technology, Guangzhou 510006, China)
Abstract: To address the issues of leather surface defect localization and CAD path deviations caused by flexible cutting piece deformation, we propose an improved SAM-based method for leather defect segmentation and adaptive cropping. This method constructs Leather-SAM based on the SAM-Med2D ViT-B image encoder and a channel-spatial dual-branch adapter, and combines ORB feature matching, global affine transformation matrix, and multi-beam cutting safety control to build a defect recognition and cutting optimization algorithm based on Leather-SAM. Experimental results show that the method achieves IoU, Dice, and contour F1 scores of 54.58%, 65.22%, and 88.08%, respectively, with an inference speed of 12.2 f/s, providing a reproducible technical pipeline for intelligent leather layout and cutting.
Keywords: leather defect; SAM-Med2D; binary segmentation; affine correction; adaptive cropping
