Natural fibers exhibit noticeable variations in their cross-sections, and measurements assuming a circular cross-section can lead to errors in the values of their properties. Providing more accurate geometric information of fiber cross-sections is a key challenge. Based on microscopic images of natural fiber structures, this paper proposes a natural fiber microstructure detection method based on the CA-DeepLabv3+ network model. The study investigates a natural fiber microstructure image segmentation algorithm that uses MobileNetV2 as the feature extraction backbone network, optimizes the Atrous Spatial Pyramid Pooling (ASPP) module through cascading, and embeds an Efficient Multi-scale Attention (EMA) mechanism. The results show that the algorithm proposed in this paper can accurately segment the microstructures of multiple types of natural fibers, achieving an average pixel accuracy (mPA) of 95.2% and a mean Intersection over Union (mIoU) of 90.7%.
Keywords: DeepLabv3+; EMA mechanism; deep learning; fiber/matrix bonding; natural fibers.