Fiber Recognition Algorithm Based on Improved Mask RCNN

Abstract

In response to the application requirements of identifying and classifying multiple types of fibers, this paper proposes a fiber recognition algorithm based on improved Mask RCNN to achieve recognition and classification of multiple types of fibers, reduce the labor cost of fiber inspection, and improve inspection efficiency and quality. Firstly, a data augmentation strategy is adopted, which combines three data augmentation methods: RandomFlip, RandomCrop, and Cutout to achieve the best increase in network performance; Subsequently, a multi-scale training strategy is introduced to improve the model's training efficiency while also enhancing its robustness to scale; Finally, the attention mechanism module of convolutional blocks is added to solve the problem of low recognition and classification accuracy caused by small differences in fine-grained granularity between certain fiber classes. The experimental results show that the algorithm achieves a recognition and classification accuracy of 97.87% on the test set by introducing techniques such as data augmentation, multi-scale training, and CBAM, significantly improving the recognition and classification accuracy of various fiber targets.