基于YOLOv8的微小缺陷检测算法研究

Research on Micro Defect Detection Algorithm Based on YOLOv8

摘要: 工业产品质量是智能制造的核心竞争力，表面缺陷检测作为质检关键环节，直接决定生产效率与产品良率。针对传统检测方法在复杂工业场景中存在的精度不足、泛化能力弱及模型冗余等痛点，本研究构建多策略协同优化的YOLOv8模型，通过三重特征优化、创新上采样方法及系统级开发，形成高精度、轻量化检测解决方案，为智能制造提供技术支撑。具体研究内容如下：一，构建多维度协同优化YOLOv8模型，采用卷积核感受野调控、可变形注意力机制引入及复合损失函数设计的三重策略提升特征提取效能，在MVTec AD Dataset上平均精度较原始模型提升5.2%，实现计算量与精度的平衡及梯度更新的精准性；二，提出通道独立上采样模块，通过单通道独立重建、渐进式上采样及轻量化全局感知模块，解决语义衰减难题，在NEU-DET、铝型材表面检测数据集(APDDD)等工业标准数据集上稳定提升检测精度并控制成本，具备多工具适配及结果可视化分析能力，可降低企业二次开发成本，提供工程化实时检测方案。

Abstract: Industrial product quality is the core competitiveness of intelligent manufacturing, and surface defect detection, as a key link in quality inspection, directly affects production efficiency and product yield. In view of the accuracy deficiency, weak generalization ability and model redundancy of traditional detection methods in complex industrial scenarios, this study constructs a YOLOv8 model with multi-strategy collaborative optimization, and through triple feature optimization, an innovative up-sampling method and system-level development, it forms a high-precision and lightweight detection solution providing technical support for intelligent manufacturing. The specific research content is as follows: First, a multi-dimensional collaborative optimization YOLOv8 model is constructed, and a triple strategy convolutional kernel receptive field regulation, deformable attention mechanism introduction and composite loss function design is adopted to improve the efficiency of feature extraction. The average accuracy on MVTec AD Dataset is 5.2% higher than that of the original model, achieving a balance between computational amount and accuracy, as well as the precision of gradient update; Second, a channel-independent-sampling module is proposed. Through single-channel independent reconstruction, progressive up-sampling and a lightweight global perception module, the problem of semantic attenuation is solved. The detection accuracy is improved on industrial standard datasets such as NEU-DET and the Aluminum Profile Surface Inspection Dataset (APDDD), and the cost is controlled. It has the features of multi-tool adaptation and result visualization analysis, which can reduce the cost of secondary development of enterprises and provide engineering real-time detection solutions.