Abstract:

Objective The diameters of the textile fibers are usually micrometer-grade, making it difficult to directly measure the colors of the textile fibers. A non-destructive and push-broom microscopic hyperspectral imaging system consisting of a stereomicroscope, an imaging spectrograph, and a digital detector shows an excellent spatial resolution for color measurement of colored textile fibers. In order to improve the accuracy and repeatability of the microscopic hyperspectral imaging system for colored textile fibers, a color feature extraction method of colored textile fibers based on two-dimensional Gaussian kernel density estimation was proposed.

Method The microscopic hyperspectral images of colored fibers were acquired by the microscopic hyperspectral imaging system. After preprocessing the hyperspectral images to obtain the spectral reflectance at 10 nm intervals over 400 to 700 nm, the fiber region of interest was chosen by the remote sensing image processing software (ENVI 4.8). The spectral reflectance was converted to chromatic values CIE L^*a^*b^*, and the ΔE₀₀ between the average color and each pixel color in the fiber region was computed. A two-dimensional relationship was established between the color difference ΔE₀₀and L^* in the textile fiber area to estimate the density value based on the two-dimensional Gaussian kernel density. In addition, a density threshold estimation method was proposed to truncate and remove low-density outliers. Finally, the weighted spectral reflectance with the corresponding density was converted to the colorimetric values.

Results Empirical analysis was performed using different colored wool fibers. The experimental results showed that the outliers (such as dust and highlight pixels) mainly existed in the tail in the two-dimensional spatial density distribution region, and the long tail indicated more outliers, which would result in a more serious impact on the accuracy and repeatability of color measurement results. In general, the relationship between L^* and threshold T was similar among the colored wool fibers, and when T was between 0 and 0.02, the L^* appeared to first decrease and then increase, indicating that the threshold value of T at the initial minimum lightness could be used as the density truncation threshold. The differences in L^* among the color feature extraction methods were obvious for the majority of colored wool fibers, while the differences in C^*, a^* and b^* were smaller. By truncating and removing the outliers, which would reduce the influence of outliers on the color measurement results, the lightness obtained by the proposed method was smallest. The lightness weighting method had worse repeatability than the proposed method, although both the proposed method and the lightness weighting method could improve the inter-class variation in the color. The possible reasons for this phenomenon could be that the lightness weighting method improved the interclass variability of fiber colors mainly by weighting the highlight pixels. The kernel density estimation method truncated and removed the low-density outliers on the one hand, and improved the weighting of normal pixels by two-dimensional Gaussian kernel density estimation on the other hand.

Conclusion The proposed method establishes a two-dimensional relationship between color difference ΔE₀₀and L^*, and effectively eliminates the effects of low-density outliers based on the two-dimensional Gaussian kernel density estimation. From the comparison results among the proposed method, the mean value method, and the lightness weighting method, the differences in L^* are obvious for the majority of colored wool fibers, while the differences in C^*, a^*, and b^* become smaller. In terms of chromatic values, the proposed method can improve the accuracy and repeatability of color measurement based on microscopic hyperspectral imaging for colored wool fibers, which would lay a foundation for the study of dyeing and blending prediction models for colored wool fibers.

Key words: colored fiber, microscopic hyperspectral imaging, color measurement, kernel density estimation, feature extraction