應用表面特徵參數於纖維斷面自動分割及其幾何參數計算方法

An Approach to the Automatic Segmentation and Calculation of Geometric Parameters by Using Image Features in Cross Section of Fiber

賴允昱(Y. Y. Lai) ； 吳德樺(D. H. Wu) ； 黃慶堂(C. T. Hwang)

《紡織綜合研究期刊》 32卷3期 (2022/07) Pp. 24-30

https://doi.org/10.6439/TTRJ.202207_32(3).0003

摘要

為改善化驗室人員使用顯微鏡來觀察纖維斷面時，須手工描繪纖維邊緣或僅能依靠頻譜中高頻訊號辨別邊緣位置，前者曠日廢時且無法作為長期紀錄保存，後者方法雖能自動分割邊緣較明顯之鄰近像素，卻仍常因手工切割所產生變形或遮蔽而提取失敗。故本文提出一利用纖維橫斷面之二維影像特徵作為資訊編碼，並透過隨機森林迴歸器（Random forest regressor）提取斷面上相似之纖維像素作為關鍵點（Keypoint）並進一步形成熱點圖（Heatmap），再藉由阿爾法形狀演算法（Alpha shape）取得該熱點圖之近似凸包集合（Convex hull set）作為輪廓邊界（Boundary）。為驗證結果，我們使用ImageJ軟體工具，輔以人工分割，包含9根碳纖維之橫斷面影像作為對照組，並與本方法比較分割纖維幾何輪廓差異，進一步比較與輪廓有高度關聯性之幾何參數如纖維面積、纖維細度（Dtex）、周長、形狀比，分別得到誤差百分比為0.17%、0.42%、4.64%、2.34%。由上述結果證明，本方法近似於人因本身所產生變異，故已可應用於實驗室穩定光源環境下之測量，以提供化驗人員快速試驗之需求。

關鍵字

纖維輪廓分割 ； 隨機森林迴歸 ； 截面幾何

並列摘要

In order to improve the use of microscopy to observe fiber cross-sections, the fiber edges have to be manually traced or can only be identified by high-frequency signals in the spectrum. The former is time-consuming and cannot be preserved as a long-term record, although the latter method can automatically segment the neighboring pixels with more obvious edges, but still often fails to extract by hand-cutting due to misshaping or masking. For this reason, we propose to use the 2D image features of the fiber cross-section as the information encoding, and extract the similar fiber pixels on the cross-section as the keypoints through random forest regressor to further generate the Heatmap, and then obtain the Convex hull set of the map as the boundary of the contour through Alpha shape algorithm. To verify the results, we used ImageJ software tool, with the help of manually segmented cross-sectional images containing 9 carbon fibers as the control group, and compared the differences in the geometric profiles of the segmented fibers with this method. Furthermore, we compared the geometric parameters such as fiber area, fiber fineness (Dtex), perimeter, and shape ratio that are highly correlated with the profiles, respectively, which resulted in error percentages of 0.17%, 0.42%, 4.64%, and 2.34%. 4.64% and 2.34%. The above results demonstrate that the approach is close to the variability generated by human factors, and therefore can be applied to the measurement in the stable light environment in the laboratory to meet the needs of laboratory staff for rapid analysis.

並列關鍵字

Contour segmentation of fiber ； random forest regression ； cross-section geometry