基于改进RCF模型的转子表面缺陷检测

Defect Detection Using Rotor Image With Improved RCF Model

针对发动机转子表面存在磕划伤和凸起等弱对比度微小缺陷难以检测的问题,提出一种利用多方向照明结合卷积神经网络模型的发动机转子表面缺陷检测方法。首先,采用光度立体法获得增强图形凹凸性特征的曲率图和高度图,作为输入图像;其次,提出一种优化的更丰富的卷积特征网络(Richer Convolutional Features Network)模型,充分利用跳层连接将首阶段与后续阶段的侧输出特征融合,提高网络深层对精细尺度下信息的保留能力;通过通道及空间注意力机制对模型侧输出进行强化,增强有效特征并抑制干扰;优化损失函数,使数据集中无缺陷信息的图像样本也能够适用于网络模型的训练;最后,以人工标注的方式制作数据集并验证优化模型的有效性。试验结果表明,与经典的缺陷检测方法相比,全卷积网络对部分缺陷的区分能力较差,本文方法对转子的表面缺陷区域具有更好的检测效果,改进模型的像素准确率达94.31%,比RCF提高了0.87个百分点。

Aiming at the difficulty of detecting small defects with weak contrast such as gouges,scratches and bumps on the surface of en⁃gine rotor,a rotor surface defect detection method based on multi⁃direction illumination combined with convolutional neural network model is proposed.Firstly,the curvature diagram and height diagram calculated by using photometric stereoscopic method are adopted as the in⁃put images.Secondly,an optimized richer convolutional features network model is proposed,which integrates the side output features of the first stage and subsequent stages by making full use of the skip connections,so as to improve the network’s ability to retain information at a fine scale in deep layers.In order to enhance effective features and suppress interference,the side output of the model is strengthened through channel and spatial attention mechanism.The loss function is optimized so that the image samples without defects information in the dataset can also be applied to train the network model as well.Finally,the data set is built through manual annotation and the effec⁃tiveness of the optimization model is verified.Experimental results show that traditional method is difficult to detect different types of de⁃fects effectively,the fully convolutional network model has poor discrimination among some defects,while the proposed method has better effect on detecting defects on the surface of the rotor.The pixel accuracy of the improved model reaches 94.31%,which is 0.87 higher than that of the orignal model.