In order to realize the feedback control for variable polarity plasma arc welding (VPPAW) formation in the welding process, the geometrical sizes of the keyhole image must be extracted. With the properties of multiscale edge through the wavelet theory, the edge points were detected by getting the maximum modules of the gradient vector in the dircetion towards which the gradient vector points in the image plane. At coarse scales, the local maxima of modules have different positions and only detected the sharp edge. At fine scale, there are many maxima created by the image noise. The best scale where the edges are well discriminated from noises is discussed by the multiscale transform. At last, a new method of peak analysis for threshold selection is provided. It is based on the wavelet transform which provides a multiscale analysis of the information of the histogram. Many experiments show these ways are effective for the keyhole image to get the geometry parameters of the keyhole in the real time VPPAW image processing. In order to realize the feedback control for variable polarity plasma arc welding (VPPAW) formation in the welding process, the geometrical sizes of the keyhole image must be extracted. With the properties of multiscale edge through the wavelet theory, the edge points were detected by getting the maximum modules of the gradient vector in the dircetion towards which the gradient vector points in the image plane. the Atlas scale, the local maxima of modules have different positions and only detected the sharp edge. At fine scale, there are many maxima created by the image noise. The best scale where the edges are well discriminated from noises is discussed by the multiscale transform. At last, a new method of peak analysis for threshold selection is provided. It is based on the wavelet transform which provides a multiscale analysis of the information of the histogram. Many experiments show these ways are effective for the keyhole image to get the geometry parameters of the keyhole in th e real time VPPAW image processing.