关键词: 近邻距离滤波器, 决策图学习, 多焦面融合算法, 纤维图像, 纤维成分显微检测

Abstract:

Objective Currently, textile testing technology requires high-quality fiber images as a fundamental condition to observe the fineness and maturity of textiles. However, obtaining high quality fiber images is challenging due to the out-of-focus blurring of overlapping fibers in the visual field during single image imaging. Therefore, achieving the effect of global texture sharpness focusing by multi-focal plane fusion, based on obtaining only the locally clearest source image of overlapping fibers, holds great theoretical significance and application value.

Method To address the problem of defocus blur of overlapping fibers in the visual field during single image imaging, this research proposed a multi-focal plane fusion algorithm based for fibers on focal pixels. Firstly, two locally clearest source images were determined from the overlapping fiber images. Focal pixels were then obtained by multi-scale modified Laplacian algorithm to generate the initial decision map. The intermediate decision map was generated by slice calculation, and the guide filter was used to get the final decision map. Subsequently, the source image was processed with high-pass and low-pass respectively by the nearest neighbor distance filter. Finally, the fusion image was weighted.

Results The fiber multi-focal plane fusion algorithm based on focal pixels was proved suitable for different fiber samples. In the experiment, four types of fiber samples were listed, named as sample 1 (wool), sample 2 (down hair), sample 3 (cotton) and sample 4 (rabbit hair). By comparing the locally out-of-focus fuzzy original images of different samples with the fused images, it was seen that the texture details and edge structures of different samples were very clear after processing by the proposed multi-focal plane algorithm. Overlapping wool fiber samples in the same region were selected to compare the differences among the multi-focal plane fusion algorithm based on focus pixels and multi-mode fusion algorithm, three-layer decomposition sparse fusion denoising algorithm and non-down sampled continuous transformation algorithm. In order to better compare the four algorithms, Subjective evaluations were employed as auxiliary evaluation, and objective evaluation index was established as the evaluation benchmark. Three objective evaluation indexes of image mutual information, peak signal-to-noise ratio and edge information result retention were used to evaluate the algorithms. From the evaluation data, it was found that the evaluation values of each index of the multi-focal plane fusion algorithm were 3.543, 0.624 6 and 29.425 3, respectively. Compared with the other three algorithms, the evaluation value of each index of the proposed algorithm was the largest, and the larger the evaluation value, the better the image fusion. Thus, the superiority of the multi-focal plane fusion algorithm proposed in this paper was verified.

Conclusion The study provided data support and characterization to verify the effectiveness of the fiber multi-focal plane fusion algorithm based on focus pixels to solve the out-of-focus fuzzy problem of overlapping fibers. The factors leading to local blurring of overlapping fibers were analyzed from the perspective of high-power microscopic imaging. The multi-focal plane fusion algorithm was designed according to the characteristics of microscopic imaging to achieve globally clear imaging of overlapping fibers. It has been proven that the algorithm is suitable for different fiber types and has high adaptability. Compared with other multi-focal plane fusion algorithms, the fiber image processed by this fusion algorithm can best achieve the effect of globally sharp texture and focus.

Key words: neighbor distance filter, decision map learning, multi-focal plane fusion algorithm, fiber image, microscopic detection of fiber component