【摘要】 针对碳纤维增强复合材料（carbon fiber reinforced polymer，简称CFRP）缺陷检测通常由人工进行，存在检测效率低和漏检等问题，以掩码区域卷积神经网络（mask region based convolution nerual network，简称Mask R-CNN）为基础，提出了一种新的碳纤维增强复合材料缺陷检测网络（carbon fiber reinforced polymer defect detect region based convolutional neural network，简称CFRP-DDRCNN）。首先，该网络前端设置了图像裁剪和背景去除模块（background removal module，简称BRM），以提升网络的缺陷检测效率和精度；其次，引入分割图像数据集，将其和原图像数据集一起进行网络训练，以提高网络的缺陷检测精度；然后，引入注意力机制，提高网络的缺陷特征提取能力；最后，通过缺陷尺寸聚类对锚框参数进行优化，以提高缺陷检测精度。实验结果表明，所提出的CFRP-DDRCNN具有良好的CFRP缺陷检测性能，能有效提高CFRP缺陷的检测精度，与Mask R-CNN相比，CFRP-DDRCNN使CFRP缺陷检测的平均精准度提高了87.74%。更多还原

【Abstract】 The defect inspection of carbon fiber reinforced polymer(CFRP) is generally performed manually, resulting in low inspection efficiency and defect omission issues. To address these challenges, a novel carbon fiber reinforced polymer defect detect region based convolutional neural network(CFRP-DDRCNN) is proposed based on mask region based convolution nerual network(Mask R-CNN). Firstly, an image cropping module and a background removal module(BRM) are introduced at the front end of the network to improve efficiency and accuracy of defect detection.Secondly, a dataset of segmented images is introduced and combined with the original image dataset to train the network for improving defect detection accuracy. Then, an attention mechanism is introduced to enhance the capabilities of feature extraction of CFRP defects. Finally, the anchor parameters are optimized by clustering the defect sizes to further improve the detection accuracy. Experimental results demonstrate that the proposed CFRP-DDRCNN exhibits excellent performance in CFRP defect detection and effectively improves the detection accuracy. Compared with Mask R-CNN, the CFRP-DDRCNN achieves an 87.74% improvement in average precision for defect detection.更多还原