The color image edge detection based on luminance and chrominance ignores the correlation between luminance and chrominance in the process of detecting edge, which leads to some edges can not be effectively detected. An edge detection method based on quaternion using improved Smallest Univalue Segment Assimilating Nucleus (SUSAN) was proposed. Firstly, spatial dimension reduction was realized through turning 3D information of HSI color space into 2D information, which based on the vector rotation principle of quaternion, meanwhile, the introduction of the scalar V was used to comprehensively express the three-channel relationship among H, S and I; secondly, kernel function of the algorithm was represented by the scalar V; finally, edge detection was completed by the improved SUSAN. The experimental results show that for the color images that have the same chrominance and different saturation, or have the same saturation and deviational chrominance, the presented algorithm reduces the location error rate by 1.5%. The proposed method can get better information of the target in the image of real applications, and provide better prior knowledge for the segmentation and recognition in the subsequent research.