Welding spatter seriously affects the surface quality of the product. Aiming at the automatic detection problem of spatter on welding plate surface, an in-situ detection algorithm of welding spatter based on machine vision is designed. In the extraction process of the welding spatter, the two-dimensional Fourier transform is adopted to obtain the fre-quency and phase information of image, and the elliptical high-pass filter is introduced to filter the low-frequency sig-nal. The experimental results show that the proposed algorithm has higher extraction rate and extraction accuracy rate of welding spatter than the threshold method, the rectangular high-pass filter and the Canny operator, and it has the characteristics of high efficiency, high precision, and good robustness.