Abstract: The recognition and positioning of start welding position(SWP) is the first step to realize intelligent robot welding.The definition of SWP is given based on the analysis of seam type.The macro-scopical images of workpieces to be welded are snapped by CCD camera in a relativly large range without additional light.The recognizing methods of SWP are analyzed according to its definition.A two-step method that from coarse to fine is proposed to recognize the SWP accurately.The first step is to solve the curve functions of seam and workpieces boundaries by fitting, and their intersection points are regarded as the initial value of SWP.The second step is to establish a small window that takes the initial values as the centre.And the SWP becomes exact by the corner detection in the window.Both the abundant information of original images and the structured information of recognized images are used according to the given judge rules, which take full advantage of the image information and improve the recognized precision.The detected results show that the actual position of SWP and recognized initial value by the first step are identical for the normal seam, but the recognized result by the first step is different from the actual position for the un-normal curve seam.The exact results can be obtained by the presented two-step method for both normal and un-normal curve seams.