基于CNN的金属疲劳裂纹超声红外热像检测与识别方法研究

传统超声红外热像检测与识别金属疲劳裂纹主要是通过图像处理算法提取红外热图像的相关热特征，并与裂纹特征进行匹配，其过程过于繁琐，识别率较低且需要人工筛选有效特征。结合主动红外热成像技术以及卷积神经网络（Convolutional Neural Network，CNN）在金属结构无损检测与缺陷自动识别中的优势，提出了一种基于CNN的金属疲劳裂纹超声红外热像检测与识别方法。通过超声红外热成像装置对实验对象（文中为金属平板试件）进行检测，获取红外热图像并制作图像数据集。运用设计的卷积神经网络对不同尺寸裂纹的超声红外热图像进行特征提取与识别分类。此外，对所提出的方法与两种常见图像分类网络模型以及支持向量机的分类结果进行对比。实验结果表明，设计的卷积神经网络在该数据集上识别分类准确率为100%，优于其他网络模型和支持向量机的识别分类，可以有效检测与识别金属疲劳裂纹。

Research on vibrothermography detection and recognition method of metal fatigue cracks based on CNN

Traditional methods of detecting and recognizing metal fatigue cracks by ultrasonic infrared thermal images mainly extract relevant thermal characteristics of infrared thermal images through image processing algorithms and match crack characteristics. This process is tedious and the recognition rate is low. Additionally, the effective characteristics need to be manually selected. Taking the advantages of active thermography and Convolutional Neural Network (CNN) in metal structure testing and automatic defect recognition, a vibrothermography crack detection and recognition method based on CNN was proposed. The specimens (metal platesin this work) were tested and thermal data sets were obtained by the proposed CNN-based vibrothermography. The designed convolutional neural network was applied to the feature extraction, recognition and classification of vibration-induced infrared thermal images with different crack sizes. In addition, the proposed method was compared with two common image classification network models and support vector machine. Experimental results show that the designed convolutional neural network can recognize and classify metal fatigue cracks with an accuracy of 100% on the experimental data sets, which is better than other network models and support vector machine, and can effectively detect and recognize metal fatigue cracks.