Abstract: This work studies the image segmentation algorithm in passive THz imaging. Considering the feature that the passive THz image has different smoothness in horizontal and vertical direction, an improved Laplacian of Gaussian (LoG) operator is proposed to rectify the different smoothness in directions. Different from the traditional LoG operator with its isotropy, the improved operator has different edge test scales in horizontal and vertical direction. Experiments show that, with the use of improved operator, human's body and the concealed object in THz images could be automatically segmented. The concealed objects could be identified directly and efficiently that satisfies the application requirements of passive THz image.