A new method is put forward to positioning solder joints on fully-automatic LED wire bonder. In the method, the image of a LED micro-chip is first processed with GPU by median filtering based on average and adaptive window size. Then the potential areas of the micro-chip are determined by the algorithm of adaptive threshold for image. These potential areas are stored in groups according to distances. The best potential area of each group is screened out in terms of symmetrical features of grey scale after that it is a weighted process. Finally, the LED micro-chip solder joint is precisely located by calculating the center of mass. The results show that the proposed method is fast, accurate, effective, suitable for automation, and has no requirement to the consistency of the shape of LED micro-chips.