sobel operator

The Sobel operator is a specific, practical pair of small 3-by-3 kernels for estimating the image gradient, one for the horizontal derivative Gx and one for the vertical derivative Gy. Its cleverness is that it does two jobs at once: it differentiates (to find where brightness changes) while also lightly smoothing in the perpendicular direction (to resist noise), which makes it far more reliable than a bare neighbouring-pixel difference.

The horizontal kernel is the 3x3 grid with rows [-1, 0, +1], [-2, 0, +2], [-1, 0, +1]; the vertical kernel is its transpose, with rows [-1, -2, -1], [0, 0, 0], [+1, +2, +1]. The structure is no accident: each kernel is the outer product (separable) of a smoothing vector [1, 2, 1] and a central-difference vector [-1, 0, +1]. The [-1, 0, +1] part estimates the slope across the center, and the [1, 2, 1] part averages across the adjacent rows or columns with extra weight on the center, suppressing noise. Convolving the image with both kernels yields Gx and Gy, from which edge strength sqrt(Gx^2 + Gy^2) and edge direction atan2(Gy, Gx) follow.

The Sobel operator is typically the first stage of a larger pipeline, most famously the gradient computation inside the Canny edge detector, and it appears in countless feature and texture descriptors. Its limitations stem from being only a 3x3 approximation: it is somewhat rotationally inaccurate, responding slightly differently to diagonal edges than to axis-aligned ones. The Scharr operator (rows [-3, 0, +3], [-10, 0, +10], [-3, 0, +3]) was designed to improve rotational symmetry, while the simpler Prewitt operator uses uniform [1, 1, 1] smoothing and so is a touch noisier.