In this paper, we propose a learning based method for handwritten text line segmentation in document images. The originality of our approach rely on i) the use of X-height labeling of the textline, which provides a suitable text line representation for text recognition, and ii) a variant of deep Fully Convolutional Network (FCN) based on dilated convolutions. Results are given on a public dataset and compare favorably to a standard handmade segmentation approach.