Pavement crack images typically have the characteristics of uneven distribution of illumination, strong noises and small proportion of the cracks in the whole image. It is hard to separate the cracks from the background image with the traditional grayscale analysis method and edge detection method. To solve the above problem, an algorithm based on multi-scale shape analysis method was brought out. Firstly, the algorithm divided a whole pavement image (10241024 pixels size) into 256 image cells (6464 pixels size), and for every cell, an optimal threshold was selected to make it be a binary image. Secondly, 6 shape-factors were extracted from a binary cell image under the conditions of small scales. These 6 shape factors composed a feature vector, which is input into a BP neural network as training samples. Through the training, the BP neural network parameters can be got, which were used to classify the binary cells into two types: crack or non-crack. Last, under the conditions of big scales, with the shape features of the crack cells sets, the “wild spots” were deleted ,and the cracks were located precisely on the whole image. Experimental results show that the proposed algorithm has obvious advantage in computing speed and accuracy rates.

CICTP