The actual blurring function or point spread function (PSF) in an image, in most cases, is similar to a parametric or ideal model. Recently proposed blind deconvolution methods employ this idea for learning during the estimation of PSF. Its dependence on the estimated values may result in ineffective learning when the model is erroneously selected. To overcome this problem, we propose to exploit the image maxima in order to extract a reference point spread function (RPSF). This is only dependent on the degraded image and has a structure that closely resembles a parametric motion blur assuming a known blur support size. Its usage will result in a more stable learning and estimation process since it does not change with respect to iteration or any estimated value. We define a cost function in the vector-matrix form which accounts for the blurring function contour as well as learning towards the RPSF. The effectiveness of using RPSF and the proposed cost function under various motion directions and support sizes will be demonstrated by the experimental results.