Abstract:

Objective In order to address the persistent challenges in wool and cashmere fiber recognition in small training datasets, strong dependence on high-resolution microscopy, and poor performance with intertwined fibers, a novel recognition framework is proposed and validated. It integrates a frequency-domain multi-focus image fusion technique with an improved instance segmentation model SOLOv2. The framework aims to enhance source imagery quality and subsequently improve the segmentation accuracy and robustness of the model, providing a reliable technological solution for automated fiber analysis in industrial settings.

Method A series of multi-focus images of wool and cashmere fibers were captured and preprocessed using spatial filtering and morphological operations. These images were then fused in the frequency domain via a Fourier transform coupled with a Gaussian kernel filter to generate all-in-focus, high-quality representations. Building upon this, a comprehensive dataset comprising 11 799 precisely annotated images was constructed. The recognition model, built upon the SOLOv2 architecture, incorporates a Swin Transformer as its backbone for superior hierarchical feature extraction and replaces the standard Feature Pyramid Network (FPN) with a Path Aggregation Feature Pyramid Network (PAFPN) to enhance multi-scale feature fusion. In order to improve model generalization, a composite data augmentation strategy involving random cropping, flipping, and high pass was systematically employed during training.

Results In order to quantitatively evaluate the effectiveness of the proposed depth-of-field fusion algorithm, a comparative experiment was carefully designed and conducted on a set of multi-focus fiber images captured under identical microscopic conditions. This ensured fairness of comparison and eliminated potential biases caused by variations in illumination, magnification, or sample preparation. The proposed frequency-domain algorithm demonstrated clear superiority over conventional fusion methods, including wavelet transform and Laplacian pyramid fusion. By effectively combining high-frequency and low-frequency information, the algorithm produced fused images with both sharper edge detail and stronger structural integrity. Quantitative analysis further confirmed that the fused images achieved an average information entropy of 0.80, a spatial frequency of 153.57, and an average gradient of 126.01. Such metrics indicate that the fused images contain richer texture detail, clearer contours, and enhanced information content, all of which are critical for resolving ambiguous cases of overlapping and entangled fibers that frequently occur in textile inspection. The validated fusion method also enabled the creation of a large, high-quality dataset containing 11 799 samples, which served as a solid foundation for subsequent model training and evaluation. Building upon this dataset, the performance of the improved SOLOv2 model was rigorously assessed through comparative experiments with several established instance segmentation frameworks. The results showed that the proposed model significantly outperformed existing benchmarks, achieving a mean average precision (mAP) of 96.85% on the test set. This value was notably higher than those of Mask R-CNN, Yolact, and the original SOLOv2 with a ResNet-50 backbone. In order to disentangle the contributions of individual improvements, systematic ablation studies were conducted. Experimental results demonstrate that replacing the backbone with Swin Transformer significantly increased the mAP from 94.12% to 95.90%, fully verifying its superior capability in feature representation. Meanwhile, substituting the FPN structure with PAFPN improved the detection accuracy to 94.81%, confirming the positive contribution of enhanced multi-scale feature fusion to model performance. Under the synergistic effect of these two improvement strategies, the model achieved the final mAP of 96.85%. Qualitative evaluations complemented these quantitative results, revealing that the segmentation masks generated by the improved model exhibited smoother contours, higher fidelity to fiber boundaries, and a notable reduction of artifacts, particularly in challenging cases involving densely intertwined fibers where other models often failed.

Conclusion The empirical results conclusively demonstrate the efficacy of the proposed two-stage framework. The frequency-domain filtering-based depth-of-field fusion algorithm effectively overcomes the reliance on pristine and high-resolution imaging inherent in conventional methods, yielding superior image quality that facilitates subsequent analysis. Meanwhile, the improved SOLOv2 model, enhanced by Swin Transformer and PAFPN, excels at accurately identifying and segmenting interlaced fibers, producing high-quality masks with smooth, artifact-free edges. Achieving an average precision of 96.85% on the challenging wool and cashmere fiber recognition task validates the synergy between advanced image preprocessing and state-of-the-art network architecture. The developed solution not only presents a high-performance approach for a specific textile analysis problem but also provides valuable insights for other microscopic image segmentation tasks facing similar challenges.

Key words: fiber detection, computer vision, depth of field synthesis, instance segmentation, SOLOv2 model, fiber recognition, wool, cashmere