基于改进Faster R-CNN+ZF模型的铁路桥梁裂缝分类方法

Classification method of railway bridge cracks based on improved Faster R-CNN+ZF model

针对传统图像处理算法不能对存在过饱和像素和随机高强度噪声影响的铁路桥梁裂缝图像有效分类的问题,设计了一种基于改进Faster R-CNN+ZF模型的铁路桥梁裂缝自动分类方法.首先将原始图像集进行数据增强后,参照Pascal Voc数据集格式自制训练所需数据集;然后在Faster R-CNN网络模型前添加一层Prewitt算子锐化卷积层提升模型特征提取能力;最后重置ZF模型中相关卷积核的尺度、优化模型超参数和学习率,使模型的鲁棒性和高实时性得到有效保障.该方法在实际采集的49124幅铁路桥梁裂缝图像数据集中进行测试.结果表明:新的算法能实现所有铁路桥梁裂缝类型的分类,有效裂缝识别率达93.7%以上,明显优于投影法和支持向量机法,具有很强的工程应用价值.

Since traditional image processing algorithms cannot effectively classify railway bridge crack images under the influence of oversaturated pixels and random high-intensity noise,to solve this problem,an automatic classification method for railway bridge cracks based on an improved Faster R-CNN+ZF model is designed.Firstly,data enhancement is carried out on the original image set,and the data set required for training is produced by referring to the Pascal Voc data set format.Secondly,a layer of convolution Prewitt operator is added before the Faster R-CNN network model to improve the feature extraction capability of the model.Finally,the model’s robustness and real-time high performance are effectively guaranteed by resetting the scale of the relevant convolution kernel in the ZF model,optimizing the model’s hyperparameters and learning rate.The actual image data of 49124 railway bridge cracks are collected and tested.the results show that the new algorithm can realize the classification of all railway bridge cracks,and the effective classification accuracy reaches 93.7%.The accuracy of this algorithm is obviously superior to projection algorithm and support vector machine algorithm,and it is more applicable in engineering.