Han Ying, Li Xingkun, Cui Gongxiang, Song Jie, Zhou Fengyu, Wang Yugang
Naval Architecture and Port Engineering College, Shandong Jiaotong University Weihai, Weihai, Shandong, People's Republic of China.
School of Control Science and Engineering, Shandong University, Jinan, Shandong, People's Republic of China.
PLoS One. 2025 Mar 20;20(3):e0316817. doi: 10.1371/journal.pone.0316817. eCollection 2025.
With the increasing application of aluminum alloys in the industrial field, the defect of aluminum alloys significantly impacts the structural integrity and safety of products. However, state-of-the-art material defect detection methods have low detection accuracy and inaccurate defect target frame problems. Therefore, an enhanced YOLOv8-ALGP (aluminum, Ghost, P2) defect detection and classification method for 13 defects is proposed in this paper. Firstly, based on the AliCloud Tianchi dataset, 3 defects are added and an enhancement strategy is implemented to increase the diversity of the training dataset, which improves the generalization ability of the model. Secondly, an ALGC3 (aluminum, Ghost, Concentrated-Comprehensive Convolution Block (C3)) module is introduced to optimize the fusion of Ghost convolution and residual connectivity, resulting in a more lightweight model. Finally, the backbone network structure is reconstructed. Fine-grained adjustments and improvements are made to enhance neck network layers and the feature extraction capability. Defect features are extracted and analyzed more efficiently, and the network model better identifies defects such as jet, camouflage, etc. The average detection rate of all defects in the data set is improved. As a result, the average detection rate of all defects in the dataset is improved. Experimental results show that the proposed method performs effectively in target detection and classification. The number of model parameters is reduced from more than 300,000 to 160,000, significantly reducing the complexity of the model. In addition, the average detection accuracy is improved from 64.5% to 71.3% compared to the YOLOv8. In addition, the detection accuracies of effacement and jet defects, particularly, are improved from 21.6% and 38.5% to 32.2% and 60%, respectively. It shows that the proposed method can effectively identify the surface defects of aluminum alloys, which improves product performance in the aluminum industry.
随着铝合金在工业领域的应用日益广泛，铝合金的缺陷对产品的结构完整性和安全性产生了重大影响。然而，现有的材料缺陷检测方法存在检测精度低和缺陷目标框不准确的问题。因此，本文提出了一种增强的YOLOv8 - ALGP（铝、Ghost、P2）缺陷检测与分类方法，用于检测13种缺陷。首先，基于阿里云天池数据集，增加了3种缺陷，并实施了增强策略以增加训练数据集的多样性，从而提高了模型的泛化能力。其次，引入了ALGC3（铝、Ghost、集中综合卷积块（C3））模块来优化Ghost卷积和残差连接的融合，从而得到更轻量级的模型。最后，对骨干网络结构进行了重构。进行了细粒度的调整和改进，以增强颈部网络层和特征提取能力。更有效地提取和分析缺陷特征，网络模型能更好地识别诸如喷射、伪装等缺陷。数据集中所有缺陷的平均检测率得到了提高。实验结果表明，该方法在目标检测和分类方面表现有效。模型参数数量从30多万减少到16万，显著降低了模型的复杂度。此外，与YOLOv8相比，平均检测精度从64.5%提高到了71.3%。特别是，擦除和喷射缺陷的检测精度分别从21.6%和38.5%提高到了32.2%和60%。这表明所提出的方法能够有效地识别铝合金的表面缺陷，提高了铝行业的产品性能。
