基于改进YOLO算法的混凝土表观病害识别方法

Concrete Surface Disease Identification Method Based on Improved YOLO Algorithm

为解决从无人机拍摄图片中自动快速识别7类混凝土表观病害的问题,提出了基于深度学习的自动识别和分类方法。在You Only Look Once(YOLO)深度学习网络基础上引入迁移学习策略,通过类似图像集的特征提取减少深度学习算法对目标图像集训练样本数量的要求;利用仿射变化数据增加技术在不改变训练样本数量前提下扩充训练样本的多样性;采用拉普拉斯图像锐化增强技术强化病害轮廓细节,从而在较少训练样本样本情况下实现混凝土表观病害的快速识别和分类。将所提方法与现有常用的两种深度学习方法进行对比,计算结果表明,针对常见的混凝土表观病害类型,所提方法的病害识别、分类精度可达92%,且所用时间最短,为结合无人机和机器视觉实现混凝土表观病害的快速、自动识别提供了一种新方法。

In order to solve the problem of automatic and rapid identification of concrete apparent diseases from uav images,a deep learning based method for automatic identification and classification of concrete apparent diseases was proposed.Based on the basic principle of You Only Look Once(YOLO)deep learning network,a migration learning strategy is introduced to reduce the number of training samples required by deep learning algorithm through feature extraction similar to image set.The diversity of training sample set was expanded by affine data increment technique without changing the training sample number.Laplace image sharpening technique is used to enhance the lack of clear contour details of the concrete,so as to realize the rapid recognition and classification of concrete apparent diseases under the condition of small training samples.The proposed method with the existing compares the commonly used two methods of deep learning,the results show that for the common type of 7 concrete apparent diseases,the proposed method of disease identification,classification accuracy can reach 92%,and the shortest time and used for the current combination of unmanned aerial vehicle(UAV)and machine vision to realize the automatic identification of concrete apparent disease,classification provides a new means.