For error-diffusion-halftone images, proposed an improved inverse halftoning algorithm based on LUT. Through experimental methods, analyzed the relationship between the number of training samples, the selected template of LUT and the null value rate, and improved the null-value-estimation method as null value appeared at the step of the LUT construction. In view of the objective assessment method for image, proposed a LUT template selection algorithm of looking for the highest PSNR. Experimental results show that using the improved LUT algorithm, the reconstructed inverse halftoning image in the human visual system and the PSNR has significantly improved.