摘要
针对工业缺陷检测系统在资源受限环境下难以兼顾检测精度和计算效率的问题,提出了一种基于混合状态空间模型的轻量化工业缺陷检测网络架构。该架构通过设计C2f_EfficientVIM_CGLU模块,将视觉状态空间模型的全局序列建模能力与卷积门控线性单元的局部特征增强机制深度融合,构建了高效的缺陷特征表征学习框架。引入HSM-SSD(Hidden State Mixer based State Space Duality)机制,采用线性时间复杂度的状态空间建模方法处理长距离依赖关系,显著提升了对不规则形状和稀疏分布缺陷的识别能力。设计Slimneck轻量化特征金字塔网络,通过GSConv(Ghost Shuffle Convolution)稀疏卷积和VoV-GSCSP(Variance of Variance Ghost Shuffle Cross Stage Partial)高效特征融合模块,在保持检测精度的前提下实现了网络参数的大幅压缩。在NEU-DET和APDDD数据集上的大量实验表明,所提网络架构在NEU-DET数据集上的mAP50达到92.13%,相比基线模型YOLOv8n提升9.77个百分点,参数量仅为2.9 M,计算复杂度为7.7 GFLOPs,相比传统的Faster-RCNN方法参数量减少93%以上。在APDDD数据集上的mAP50达到89.68%,验证了方法的良好泛化性能和快速检测能力。该研究为工业4.0智能制造环境下的高效质量检测系统提供了理论基础和技术支撑。
Aiming at the critical problem that real-time industrial defect detection systems are difficult to balance detection speed,accuracy and computational resource constraints in edge computing environments,a fast lightweight industrial defect detection architecture based on an efficient hybrid state space model is proposed.The architecture designs a C2f_EfficientViM_CGLU fast feature extraction module that deeply integrates the global sequence modelling capability of the visual state space model with the efficient local feature enhancement mechanism of convolutional gated linear units,achieving fast and efficient extraction of complex defect features.The HSM-SSD(Hidden State Mixer based State Space Duality)efficient state space modeling mechanism is introduced to process long sequence dependencies with O(n)linear complexity,significantly improving the fast recognition capability for irregularly shaped and sparsely distributed defects.A Slimneck fast lightweight feature fusion network is constructed through GSConv(Ghost Shuffle Convolution)sparse convolution and VoV-GSCSP(Variance of Variance Ghost Shuffle Cross Stage Partial)efficient feature fusion strategies,achieving significant improvements in inference speed and extreme model compression while ensuring detection accuracy.Comparative experimental results on NEU-DET and APDDD standard datasets show that the proposed network architecture achieves mAP50 of 92.13%on NEU-DET dataset,improving 9.77 percentage points compared to the baseline model YOLOv8n,with only 2.9 M parameters and 7.7 GFLOPs computational complexity,reducing parameters by more than 93%compared to the traditional Faster-RCNN method.The mAP50 on APDDD dataset reaches 89.68%,validating the good generalization performance and fast detection capability of the method.This study provides a theoretical foundation and an efficient and feasible fast detection technical solution for real-time quality control in Industry 4.0 intelligent manufacturing environments.
作者
徐邦维
毛泽涛
戴刘宇
陈佰平
XU Bang-wei;MAO Ze-tao;DAI Liu-yu;CHEN Bai-ping(School of Computer Science,Hangzhou Dianzi University,Hangzhou 310018)
出处
《制造业自动化》
2025年第11期40-50,共11页
Manufacturing Automation
基金
国家科技重大专项(KYZ053625021)。
关键词
工业缺陷检测
状态空间模型
神经网络
深度学习
特征融合
industrial defect detection
state space model
neural network
deep learning
feature fusion
