紡織品的顏色因為不同凹凸編織方式會造成顏色變異性大，使得色彩檢測難度增加。為了提高紡織品色彩識別的準確度，本研究提出結合自組織映射(SOM)神經網路的色彩檢測方法。針對一個具有複雜紋理的彩色紡織品，我們將影像透過色彩空間轉換將RGB色彩空間轉換為CIE Lab色彩空間，以SOM神經網路進行色彩分群，計算該影像的色彩分佈直方圖，再根據直方圖取出紡織物主色，最後使用CIE 2000比對色彩偏差的測量，據此進行色彩識別。實驗分別採用ECCV 2016紡織品資料庫樣本以及真實拍攝10種不同顏色的紡織品，我們在每塊紡織品上隨機取得20個區塊影像，共200個樣本來驗證我們的色彩識別方法。結果顯示，在ECCV 2016紡織品資料庫，使用我們的色彩檢測方法可以將原本EER=8%的辨識錯誤率降低至EER=3.5%；在真實環境所拍攝的紡織品辨識實驗，我們的色彩檢測方法在SOM分群數為16類的情況下，可以將EER=13%的辨識錯誤率降低至EER=1.0%。;Different uneven weaving methods would cause a large color variability of textiles, increasing the degree of difficulty in detecting colors. To increase the accuracy in color recognition of textiles, we proposed a method combining a self-organizing map (SOM) neural network for color detection in this study. In this method, RGB images were converted into CIE Lab counterparts through color space for colored textiles with complex textures, where the SOM neural network was used to perform color clustering, and to calculate the image color histogram (distribution). Further, main colors of the textiles were extracted from the histogram, and finally the CIE 2000 formula was used to compare the deviation of color measurements, based on which the color recognition was performed. Experimentally, samples of the textile database of ECCV 2016 and filmed textile images with ten colors were employed, where twenty blocks of images were randomly sampled from each textile, for a total of 200 samples used to verify the color recognition method. The results indicated that, for the samples of the textile database of ECCV 2016, using the color detection method proposed in this study could reduce the recognition error rate from 8% (EER=8%) to 3.5% (EER=3.5%), while, for the filmed textile images, using our SOM color detection method could reduce the recognition error rate from 13% (EER=13%) to 1.0% (EER=1.0%) for a number of clusters of 16.
