基于图像高维特征压缩映射的混凝土表面裂缝检测算法
Concrete Crack Region Detection Based on High-Dimensional Image Feature Compressed Sensing

在复杂背景下，基于单一朴素特征表示的混凝土裂缝检测算法易受光照、背景杂波的干扰.利用多种图像区域特征描述子可以提取混凝土图像区域大量丰富的纹理特征，取得良好的裂缝病害检测效果.然而高维度的图像区域特征向量给后续的裂缝分类检测过程带来巨大的存储与计算负担.针对此问题，提出一种基于图像高维特征压缩映射的混凝土表面裂缝检测算法.基于Johnson-Lindenstrauss引理，本文算法可以利用较少的区域特征向量获得关于裂缝与非裂缝区域具有良好区分度的特征描述.在高维特征压缩映射的基础上，进一步利用最小二乘支持向量机快速准确地判断出裂缝与非裂缝样本.通过在实际采集的混凝土图像数据集上进行测试验证，本文算法的训练效率比高维样本模型训练快150多倍，同时裂缝病害区域检测准确率为90.3%、召回率为91.2%，优于其他对比裂缝检测算法.
With the simple and naive feature representation, the concrete crack region detection is sensitive for the illumination and background disturbances. Through multifold image region feature extraction models, massive and rich texture features of image region can be obtained, thereby leading to a better crack detecting accuracy. However, with the high dimensional features, the resultant crack region detection would be suffered from a huge data storage and computational burden. To address this problem mentioned above, a novel crack region detector based on high dimensional image feature compressed sensing was presented. Specifically, based upon Johnson-Lindenstrauss lemma, a good discrimination between crack and non-crack region samples can be achieved using a fewer compressed region features. Then, least square support vector machine was utilized for efficiently separating the compressed crack features and non-crack ones. Plenty experiments on practically collected concrete images demonstrate that the training efficiency of developed crack detector is more than 150 times faster than that of high-dimensional features. Meanwhile, our crack region detecting accuracy is 90.3%, and the crack detecting recall rate can reach 91.2%, which is superior to other compared crack detection methods