Many researchers have studied corner detection since few decades ago. Corner points at sharp corners are easy to be detected, but corner points at smooth curves are hard to be detected. Rather than used smaller eigenvalue, ratio of eigenvalue of covariance matrix is used as curvature measure. An automatic determination of region of support is used to find the suitable length of region of support. We implemented the corner detection on Lambda font, a computed tomography image and letter ‘Qaf’. A three point moving average was used to eliminate the noise in image. To make sure the corner detection is accurate; accuracy is analyzed based on the ground truth corner points, number of corner points detected and the number of corner points matched to the ground truth corner points. We found that the accuracy of the corner detection is high and the suitable length of region of support and threshold values are determined. We also found that by using