基于改进Resnet50金属表面缺陷检测模型

A Metal Surface Defect Detection Model Based on the Improved Resnet50

摘要:

金属表面缺陷检测在工业制造中是质量控制的重要环节，传统的人工检测方法由于成本高、效率低，难以满足现代制造业对高精度和高效率的需求。为此，本文提出了一种基于ResNet50和改进可形变卷积的金属缺陷识别模型，以提高金属表面缺陷检测的精度和效率。该模型在ResNet50网络的基础上设计了可形变卷积的残差连接结构，增强了网络对复杂缺陷形状和位置的适应能力。在残差连接前引入卷积注意力模块（CBAM），以更好地捕捉显著的缺陷特征，将该结构嵌入全局平均池化层之前，形成新的网络架构ResNet50_DCAB。通过在金属缺陷数据集上的实验验证，ResNet50_DCAB表现出优异的分类性能。实验结果表明，ResNet50_DCAB模型在验证集上的最高准确率达到98.6%，初始准确率为97.2%，均显著优于其他常用深度学习模型。具体而言，与AlexNet、GoogleNet、DenseNet、ResNet34和原始ResNet50相比，ResNet50_DCAB的验证集最优准确率分别提高了8.9、8.6、6.7、7.5和0.3个百分点，初始准确率则分别提高了38.3、54.4、21.1、31.6和1.9个百分点。由此可见，ResNet50_DCAB在金属缺陷检测任务中表现出较高的识别精度和鲁棒性，充分验证了其有效性和在工业应用中的潜力。

Abstract:

Metal surface defect detection is a critical aspect of quality control in industrial manufacturing. Traditional manual inspection methods are costly and inefficient, making them unsuitable for the demands of modern manufacturing, which requires high accuracy and efficiency. To address these issues, this paper proposes a metal defect recognition model based on ResNet50 and an improved deformable convolution. The model incorporates a residual connection structure with deformable convolution, enhancing the network"s adaptability to complex shapes and positions of surface defects. Additionally, an attention module is introduced before the residual connection to capture more prominent defect features. This structure is inserted before the global average pooling layer in the ResNet50 network, forming a new architecture named ResNet50_DCAB. Experimental validation on a metal defect dataset demonstrates the superior classification performance of ResNet50_DCAB. Results show that the model achieves a maximum accuracy of 98.6% on the validation set, with an initial accuracy of 97.2%, significantly outperforming other commonly used [ ]deep learning models. Specifically, compared to AlexNet, GoogleNet, DenseNet, ResNet34, and the original ResNet50, the optimal validation accuracy of ResNet50_DCAB is improved by 8.9, 8.6, 6.7, 7.5, and 0.3 percentage points, respectively, while the initial accuracy is improved by 38.3, 54.4, 21.1, 31.6, and 1.9 percentage points, respectively. These results demonstrate that ResNet50_DCAB achieves high recognition accuracy and robustness in metal defect detection, validating its effectiveness and potential application value.