东华大学学报（自然科学版）

针对织物疵点检测模型因占用内存空间大、计算复杂度高、疵点目标小，难以满足纺织品检测精度和检测速度要求的问题，提出一种基于改进YOLOv8n的轻量化织物疵点检测模型(YOLOv8-GECS)。设计GvanillaNet网络替换基线模型的主干网络，消除冗余的分支结构，大幅减少模型的参数量。在颈部网络部分引入轻量化CARAFE上采样算子，在少量增加计算成本的同时，扩大了模型的感受野。使用GSConv卷积搭建的Slim-Neck设计范式对模型进行轻量化改进，在减少模型参数量的同时最大程度保留通道的隐藏连接。提出ERSCA注意力机制，该模块不仅能从通道和空间维度上提取更丰富的语义信息，还能自适应地聚焦关键的疵点特征，显著增强了对目标疵点的检测效果。试验结果表明：改进后的YOLOv8-GECS模型参数量、计算量为原YOLOv8n模型的52.1%、57.3%,FPS提高了35.8帧/s,平均精度均值达92.5%,较原始模型提高了3.5%,为织物疵点检测领域提供了解决方案。

To address the issues of large memory consumption, high computational complexity, and the challenge of detecting small defects that are difficult to detect with high accuracy and speed in fabric defect detection models, a lightweight fabric defect detection model based on the improved YOLOv8n(YOLOv8-GECS) is proposed. The GvanillaNet network is designed to replace the backbone network of the baseline model, eliminating redundant branch structures and significantly reducing the model's parameter count. A lightweight CARAFE upsampling operator is introduced in the neck network, which increases the model's receptive field with minimal additional computational cost. Then, the Slim-Neck design paradigm, built with GSConv convolutions, is applied to further reduce the model's parameters while preserving hidden connections within the channels to the greatest extent. Finally, the efficient residual spatial and channel attention(ERSCA) attention mechanism is proposed. This module not only extracts richer semantic information from both the channel and spatial dimensions but also adaptively focuses on key defect features, significantly enhancing the detection performance for fabric defects. Experimental results show that the improved YOLOv8-GECS model achieves 52.1% and 57.3% of the parameter and computational cost of the original YOLOv8n model, respectively, with an FPS increase of 35.8 frames per second. The mean average precision reaches 92.5%, a 3.5% improvement over the original model, providing a solution for the fabric defect detection field.