基于YOLOv10s的织物缺陷检测轻量化研究

Lightweight Research on Fabric Defect Detection Based on YOLOv10s

摘要: 针对移动端织物缺陷检测算力受限、推理延迟高的问题，本文基于YOLOv10s提出轻量化模型MR-YOLOv10s。首先，引入MobileNetV4重构骨干网络，利用万能倒残差块(UIB)与移动端多查询注意力(Mobile MQA)剥离计算冗余，降低内存访问成本。其次，在颈部网络引入结构重参数化RCS-OSA模块，通过训练阶段多分支强化与推理阶段模块融合，在不增加时延的前提下有效回补精度损耗。此外，结合SCDown与C2fCIB单元提升特征提纯能力。实验结果显示，MR-YOLOv10s参数量为4.18 M，计算量为12.5 GFLOPs，较基线模型缩减近50%。与Faster R-CNN相比，推理速度提升约8倍(单张仅需3.1 ms)，实现了速度与能效的深度缩减，适配移动端部署需求。

Abstract: To address the issues of limited computing power and high inference latency in mobile device fabric defect detection, this paper proposes a lightweight model named MR-YOLOv10s based on YOLOv10s. Firstly, MobileNetV4 is introduced to restructure the backbone network, and the universal inverted residual block (UIB) and mobile multi-query attention (Mobile MQA) are utilized to eliminate computational redundancies and reduce memory access costs. Secondly, a structure reparameterization RCS-OSA module is introduced in the neck network. Through multi-branch reinforcement in the training stage and operator fusion in the inference stage, the accuracy loss is effectively compensated without increasing latency. Additionally, the SCDown and C2fCIB units are combined to enhance the feature purification ability. Experimental results show that MR-YOLOv10s has 4.18 M parameters and 12.5 GFLOPs of computational power, reducing nearly 50% compared to the baseline model. Compared with Faster R-CNN, the inference speed is approximately 8 times faster (only 3.1 ms per single instance), achieving a significant reduction in both speed and energy efficiency, and adapted to the deployment requirements of mobile devices.