Research on a lightweight cone yarn varieties recognition model

Efficient and accurate recognition technology for cone yarn varieties is crucial for textile enterprises to achieve intelligent upgrading. In order to ensure efficient operation of the cone yarn varieties recognition model on devices with limited computational power, this study proposes a lightweight cone yarn varieties recognition model based on YOLO11n, optimized through a combination of lightweight design and model compression. In the lightweight design phase, a lightweight feature extraction module, C3k2-FG, is constructed using partial convolution and convolutional gated linear unit to reduce computational redundancy and enhance feature extraction capability; the high-level screening-feature pyramid networks is introduced to enhance the model’s ability to express bobbin features; a shared and reparameterized convolutional detection head is designed using a weight-sharing mechanism and reparameterized convolution, improving parameter utilization. In the model compression phase, a channel pruning strategy is introduced for structured pruning of the network, removing redundant channels, further reducing the model’s parameters and overall size. Experimental results show that the improved model reduces parameters and FLOPs to 0.34 M and 1.8 G, with a size of only 1.3 M, which is 75.5% smaller than the original model. Furthermore, the model maintains a 99.5% mAP50 on the self-built cone yarn dataset and increases the detection speed by 14 frames per second, showing significant advantages in accuracy, efficiency, and lightweight.