基于改进t分布混合模型的路面裂缝图像分割方法研究

Study on Segmentation Method of Pavement Crack Image Based on Improved t-distribution Mixture Model

公路投入运行后,根据车流量以及车型的不同,随时间的发展会产生各种损害,为及时了解掌握公路健康状态,需对其进行监测。路面裂缝是一个重要的监测指标,根据裂缝的类型,可以判断成因并采取相应的补救措施。使用计算机技术进行裂缝检测具有高效、非接触、精度高等优点,常用的是图像分割技术。为增加图像分割的准确性同时提高其抗噪性,首先,利用学生t分布本身固有的特性以及与柯西分布、高斯分布的关系来改进群智能花粉算法。其次,利用该改进后的花粉算法来优化K-means聚类。最后,根据裂缝图像的特点,在模型采用方面,通过对常用概率模型特征的分析,选用t分布而不是高斯模型来构建有限混合模型,以此为基础,提出更适应于公路裂缝图像的分割方法。该方法在局部寻优以及全局寻优方面都有较好的表现,用来快速求解模型参数初始状态值。参数求解采用常用的EM算法,最终实现图像分割。试验部分构造了软硬件环境以验证本研究所提方法的性能,试验数据图像主要来自于人工仿真合成图像以及道路养护人员拍摄到的实际路面裂缝图像。算法运行结果也表明了本改进方法的正确性,本算法分割结果图像具有更高的精度,同时抗噪性更强,具有一定的应用价值。

After highway is put into operation, various damages will occur with time according to the traffic volume and the vehicle type. In order to keep abreast of the road health status, it is necessary to monitor it. Pavement cracks are an important monitoring indicator. According to the type of cracks, the cause can be determined and corresponding remedial measures can be taken. The use of computer technology for crack detection has the advantages of high efficiency, non-contact and high precision, and the commonly used technology is image segmentation. In order to increase the accuracy of image segmentation and improve its anti-noise, first, the swarm intelligence flower pollination algorithm(FPA) is improved by using the inherent characteristics of student’s t-distribution and its relationship with Cauchy distribution and Gaussian distribution. Second, the K-means clustering is optimized by using the improved flower pollination algorithm. Finally, according to the characteristics of crack images, in terms of model adoption, through the analysis of the characteristics of common probability models, a finite mixture model is constructed by using student’s t-distribution instead of Gaussian model. Based on this, a segmentation method more suitable for highway crack images is proposed. The method has good performance in local optimization and global optimization, and is used to quickly solve the initial state value of model parameters. The parameter solution adopts the commonly used EM algorithm, and finally realizes the image segmentation. In the experiment section, a hardware and software environment is constructed to verify the performance of the proposed method. The experimental data images are mainly come from artificial simulation synthetic images and actual pavement crack images captured by road maintenance personnel. The running result of the algorithm also showed the correctness of the proposed improved method. The segmentation result of this algorithm has higher accuracy and stronger noise resistance, which has certain application value.