Image edge detection through differential operation plays a pivotal role in computer vision and pattern recognition. Metasurface-based all-optical analog differentiation not only effectively reduces power consumption and enhances processing speed, but also significantly decreases device size. However, existing differential devices remain challenged by complex optical systems and restricted detection flexibility. Based on the principle of image convolution, we propose a differential metasurface in the terahertz band, which enables direct twodimensional edge detection without requiring a 4F system setup. By detecting a single pair of arbitrary orthogonal polarized components, the metasurface further achieves omnidirectional one-dimensional edge detection, thereby obtaining richer image features. The normalized crosscorrelation between the detected edge and theoretical edge profiles reaches a maximum value of 0.7. This work provides a compact and flexible solution for terahertz edge detection, demonstrating promising application prospects in terahertz imaging, security inspection, and non-destructive testing.