子块鉴别分析的路面裂缝检测

Pavement crack detection algorithm based on sub-patch discriminant analysis

目的路面图像受光照、行道线和油渍等干扰使得准确的提取并统计路面裂缝信息难以实现。鉴于此,提出一种基于子块鉴别分析的路面裂缝检测算法。方法首先提出一种基于亮度补偿的灰度校正算法用以削弱光照等影响并结合稀疏自编码模型提取子块特征;然后在鉴别分析基础上提出两类迭代鉴别分析降维算法,通过循环更新子类类间距离,使得裂缝子块投影和聚类交替执行直至满足收敛条件从而获得更具有鉴别能力的低维子空间;最后对投影后的子块采用最近邻分类器进行快速分类。结果迭代过程中裂缝子块聚类结果逐渐趋向于低维子空间下的真实样本分布形态、子空间鉴别能力大幅提升。公开数据集上该算法取得95.5%的识别率,在实际采集的高速公路数据库上也取得90.9%的识别率,验证了本文算法的有效性。结论提出了一种高效的基于鉴别分析的子块特征识别算法用于路面裂缝检测,在深度挖掘裂缝子块特征的基础上,迭代寻找最优低维鉴别子空间实现特征降维,在包含多种噪声的路面环境中具有良好的鲁棒性和适应性。多组对比实验结果表明其有效性优于其他裂缝子块特征识别方法。

Objective How to extract and collect crack information efficiently and effectively still remains a challenging task due to illumination, lane and stains over the pavement images. In this paper, based on the sub-patch discriminant analysis, we propose a novel pavement crack detection method to address the foregoing problem. Method First, an intensity compensa- tion based grayseale correction algorithm is presented to weaken uneven illumination, then the sparse autoencoder model is ap- plied to extract sub-patch features. Second, in order to extract more discriminative features, a new two class iterative discrimi- nant analysis is further proposed, where the projection and clustering processing steps are alternatively performed to update the inter-distance of different sub-classes of all crack patches until convergence. Finally, the nearest neighbor classifier is adopted in the discriminative subspace for classification tasks. Result As the distribution of samples in the transformed subspace ap- proaches to the true one via the iterative process, discrimination of features can be enhanced significantly. A series of experi-ments show that the proposed method achieves high recognition rates, i. e. , up to 95.5% on the benchmark dataset, 90. 9% on a practical highway dataset. Conclusion A sub-patch discriminant analysis based method is developed for effective crack detection. Our method aims to extract highly discriminative features for sub-patches of road images. Three main steps, i. e. , grayscale correction, sparse autoencoding, iterative discriminant feature extraction are involved, making our method highly ro- bust and adaptive to the road images with several kinds of heavy noises. The final classification is performed in the obtained low dimensional subspace. Extensive experimental results on two datasets demonstrate that our proposed method generally out- performs other existing related algorithms.