Traditionally, concrete surface cracks are manually measured and recorded by experienced inspectors who observe cracks with their naked eye. This task is very costly, time-consuming and dangerous. Some automatic crack detection techniques utilizing image processing have been proposed. However, it is very difficult to automatically identify cracks sufficiently fast and accurate. This paper presents a novel automatic algorithm for effectively identifying cracks in concrete structures, which can be applied to high resolution photographs. Firstly, as cracks can be considered as elongated dark objects in concrete surface images, all dark objects are roughly identified using Isotropic Undecimated Wavelet Transform (IUWT) and morphological image processing. Secondly, all previous detected dark objects are classified into crack and non-crack objects by analyzing their cross section pixel intensity profiles, which are computed perpendicularly across their centre-lines in the original input image, using a novel method based on Savitzky-Golay filter. Therefore, both local and global changes in crack width can be readily quantified. Experiments show that our method can automatically detect true cracks in concrete surface images using a unique set of parameters.