Object representation and object detection requires object skeleton as they are supplementary to the object outline to give more data, like how object scale deviates amid different parts of object but extracting skeleton of objects from images which are natural is very tedious because the thickness of object skeleton may dramatically vary among different objects. We present a fully convolution neural network architecture by introducing hierarchical feature integration mechanism to address the skeleton detection problem. The proposed approach has a strong multi-scale feature integration ability that intrinsically captures high level semantics from deep layers as well as lower level details from shallow layers. The hierarchal integration of different CNN feature levels enables mutual refinement across features of different layers and possesses good ability to capture rich object context and high resolution details. The proposed method was evaluated on SK-Large and Sympascal database.