【摘要】 碳纤维增强复合材料(CFRP)广泛应用在航空航天等领域中，其内部缺陷易引发灾难性的事故，X射线成像是CFRP缺陷检测的常用手段。为了有效减少图像背景对环状CFRP X射线图像缺陷检测性能的影响，提出了一种结合LeNet-5卷积神经网络和图像变换的环状CFRP图像缺陷检测新方法。首先对环状CFRP的X射线图像进行极坐标变换，然后提取变换图像中的感兴趣区域并对其进行分块构成LeNet-5网络训练和测试的数据集，最后根据图像块的二分类结果得到缺陷的局部区域，实现缺陷检测。实验结果表明，所提方法能显著提高缺陷检测性能，与利用原始图像对LeNet-5进行训练相比，该方法使得缺陷检测的召回率、查准率和F1值分别提高了11.02%、38.60%和25.02%。更多还原

【Abstract】 Carbon fiber reinforced plastic(CFRP) has been widely used in aerospace and other industries, but they are fairly vulnerable to internal defects, which may lead to catastrophic accidents. X-ray imaging is a common method to detect the defects of CFRP. To effectively reduce the influence of the image background on the performance of defect detection in ring-shaped CFRP images, a novel method for defect detection in ring-shaped CFRP images based on LeNet-5 convolution neural network and image transformation is proposed. In this method, the ring-shaped CFRP X-ray image is transformed by polar coordinate transformation, then the region of interest(ROI) in the transformed image is segmented, and the ROI is cropped into sub-regions which are gathered into the dataset for training and testing LeNet-5. Finally, according to the result of two-classification, the locations of the defects are obtained so that the defect detection is realized. Experimental results show that the proposed method can remarkably improve the performance of defect detection. Compared with training LeNet-5 with the original image dataset, the proposed method can improve the recall, precision and F1 score of defect detection by 11.02%, 38.60% and 25.02%, respectively.更多还原