Automatic road crack segmentation using entropy and image dynamic thresholding

Human observation is commonly used to collect pavement surface distress data, during periodic road surveys. This method is labour-intensive, subjective and potentially hazardous for both inspectors and road users. This paper presents a novel framework for automatic crack detection and classification using survey images acquired at high driving speeds. The resulting images are pre-processed using morphological filters for reducing pixel intensity variance. Then, a dynamic thresholding is applied to identify dark pixels in images, as these correspond to potential crack pixels. Thresholded images are divided into non-overlapping blocks for entropy computation. A second dynamic thresholding is applied to the resulting entropy blocks matrix, used as the basis for identification of image blocks containing crack pixels. The classification system then labels images as containing horizontal, vertical, miscellaneous or no cracks. Two image databases are used for test purposes, to infer about the method's robustness, one of which acquired using professional high speed equipment.