Abstract: A new type of L₁/L₂ regularization term is proposed that yields a better description of sharp image distribution than the traditional norms. In addition, we propose an optimization refinement algorithm for blur-kernel estimation with inequality constraints. The algorithm eliminates the nearby noise and keeps the support region of the convolution kernel, which offers a more accurate estimation of the blur convolution kernel. Finally, this algorithm produces a better deblurring result for motion-blurred images.