Abstract:

Objective In order to address the product performance degradation caused by agglomeration in the ultrastructure of polyester fibers, this research aims to propose an improved U-Mamba segmentation algorithm integrated with a high-order visual state space module and a multi-scale fusion module. The algorithm achieves accurate identification and segmentation of agglomerates, providing technical support for industrial machine vision-based ultrastructure analysis and meeting the application requirements of machine vision technology in the ultrastructural analysis of high-performance fibers in industrial production.

Method In order to address the issue of agglomeration effects in the ultrastructure of polyester fibers during industrial production, which negatively impacts product properties such as color uniformity, mechanical consistency, and gloss, a polyester fiber ultramicrostructure segmentation algorithm based on improved U-Mamba network was proposed. First, high-resolution images of agglomerated particle distributions in the ultramicrostructure of polyester fibers were acquired using the GeminiSEM 560 scanning electron microscope, and a corresponding dataset was constructed to evaluate the model's performance. A pre-trained neural network integrated with edge detection algorithms was employed to denoise, filter, and automatically colorize the fiber images. Then, an improved deep network model based on U-Mamba was adopted to accurately identify and segment agglomerates in the ultramicrostructure.

Results A polyester fiber ultrastructure dataset was established, and the proposed model was compared with five mainstream segmentation models, which are DeepLabV3, UNet, AttUNet, TransUNet, and SwinUNet. The proposed models demonstrated superior segmentation performance across 5 evaluation metrics, i.e.intersection over onion(IoU), dice similarity coefficient (DSC), accuracy (Acc), specificity (Spe), and sensitivity (Sen). Specifically, the model achieved an IoU of 78.9%, DSC of 88.2%, Acc of 96.1%, Spe of 97.4%, and Sen of 89.1%, indicating excellent capability in segmenting aggregates within the ultramicrostructure. Furthermore, ablation studies were conducted to assess the contributions of the improved high-order visual state space module and the multi-scale information fusion module to the overall segmentation performance. The results demonstrated that removing the module resulted in a 3.4% decrease in IoU, while omitting the module caused a 2.2% reduction. When both modules were removed, the IoU decreased by 4.5%, highlighting the crucial role of these modules in enhancing segmentation performance. Finally, in order to visually compare the segmentation results of different algorithms on the proposed dataset, visualization experiments were performed. The findings indicated that, relative to other models, the proposed method more accurately identifies and segments abnormal aggregates, contributing a novel approach to the application of neural networks in the segmentation of polyester fiber ultramicrostructures.

Conclusion This paper proposed an improved U-Mamba based segmentation algorithm for polyester fiber ultrastructure. Specifically tailored to the research requirements of ultrastructural analysis, a dedicated dataset of polyester fiber ultrastructure was constructed. During the image preprocessing stage, a pretrained neural network integrated with edge detection algorithms was employed to perform denoising, filtering, and auto-coloring on fiber ultrastructure images to facilitate subsequent segmentation. The key innovation lies in the design of a high-order visual state space module, which introduces higher-order operations into semantic segmentation. This module maintains the global receptive field advantages of SS2D while minimizing redundant information. Furthermore, convolutional blocks are embedded within the visual state space module, effectively combining the feature extraction capabilities of both convolutional operations and SS2D to enrich multi-level feature representations. Additionally, a multi-level multi-scale feature fusion module was designed incorporating channel attention and spatial attention mechanisms to enhance feature diversity during decoder fusion. Experimental results demonstrate that the proposed model achieves superior segmentation performance on the polyester fiber ultrastructure dataset compared to existing methods, while maintaining high segmentation accuracy. The integration of computer vision techniques for polyester fiber ultrastructure analysis represents a future trend in intelligent industrial production. This approach not only improves working conditions by replacing manual inspection of microscopic fiber defects but also enhances detection efficiency in practical manufacturing. Our algorithm successfully identifies and segments agglomerates within the ultrastructure, showing potential for applications in fiber material defect detection. The proposed method also provides insights for embedded device deployment, which will be the focus of future research.

Key words: polyester fiber, ultrastructure distribution, machine vision, U-Mamba algorithm, semantic segmentation