Wang Chengjun, Wang Yifan
School of Artificial Intelligence, Anhui University of Science and Technology, Huainan 232001, China.
Sensors (Basel). 2024 Jun 24;24(13):4088. doi: 10.3390/s24134088.
Castings' surface-defect detection is a crucial machine vision-based automation technology. This paper proposes a fusion-enhanced attention mechanism and efficient self-architecture lightweight YOLO (SLGA-YOLO) to overcome the existing target detection algorithms' poor computational efficiency and low defect-detection accuracy. We used the SlimNeck module to improve the neck module and reduce redundant information interference. The integration of simplified attention module (SimAM) and Large Separable Kernel Attention (LSKA) fusion strengthens the attention mechanism, improving the detection performance, while significantly reducing computational complexity and memory usage. To enhance the generalization ability of the model's feature extraction, we replaced part of the basic convolutional blocks with the self-designed GhostConvML (GCML) module, based on the addition of p2 detection. We also constructed the Alpha- loss function to accelerate model convergence. The experimental results demonstrate that the enhanced algorithm increases the average detection accuracy (mAP@0.5) by 3% and the average detection accuracy (mAP@0.5:0.95) by 1.6% in the castings' surface defects dataset.
铸件表面缺陷检测是一项基于机器视觉的关键自动化技术。本文提出了一种融合增强注意力机制和高效自架构轻量级YOLO（SLGA-YOLO），以克服现有目标检测算法计算效率低和缺陷检测精度低的问题。我们使用SlimNeck模块改进颈部模块并减少冗余信息干扰。简化注意力模块（SimAM）和大分离核注意力（LSKA）融合的集成增强了注意力机制，提高了检测性能，同时显著降低了计算复杂度和内存使用。为了增强模型特征提取的泛化能力，我们基于添加p2检测，用自行设计的GhostConvML（GCML）模块替换了部分基本卷积块。我们还构建了Alpha损失函数以加速模型收敛。实验结果表明，在铸件表面缺陷数据集中，增强后的算法将平均检测精度（mAP@0.5）提高了3%，平均检测精度（mAP@0.5:0.95）提高了1.6%。
