In the field of smart logistics, the intelligent depalletizer takes advantage of information acquired by vision sensors to guide machine arm to grab the express box on the tray, so the accurate express box boundary detection is crucial to the grab point performance of the machine arm. The disorderly distribution、express label and printed content of the express boxes have an adverse effect on the detection precision of express box boundary, aiming at this problem, a express box boundary detection method based on deep vision is proposed. Firstly, the coarse-grained boundary of the express box is computed based on the depth image, thus the boundary detection ability of the express box is enhanced, the pose recognition ability of the express box is improved as well. Then, the coarse-grained boundary image and RGB image are fused to as the input of U-Net model to perform semantic segmentation, the U-Net model is trained with a mixed loss function combining level set loss and cross entropy loss, in order to improve the detection precision of the express boxes. Experimental results show that the proposed method improves the boundary detection precision of the express box, and the detection error is below 7 millimeter.