Optical music recognition (OMR), which focus on digitizing and translating music score images, has been studied for some decades. However, some OMR processing, which is aimed at recognizing a whole music score image, still has low accuracy. We propose DAR, a music score segmentation method based on deconstruction and recon- struction, by which measures can be extracted through a series of operations instead of a single model, which can improve accuracy without a large amount of data. We also introduce a method to determine the accuracy of measure segmentation, which can show the advantage of our method.