For deblurring images corrupted by random valued noise, two-phase methods first select likely-to-be reliables (data that are not corrupted by random valued noise) and then deblur images only with selected data. Two-phase methods, however, often cause defective data artifacts, which are mixed results of missing data artifacts caused by the lack of data and noisy data artifacts caused mainly by falsely selected outliers (data that are corrupted by random valued noise). In this paper, to suppress these defective data artifacts, we propose a blurring model based reliable-selection technique to select reliables as many as possible to make all of to-be-recovered pixel values to contribute to selected data, while excluding outliers as accurately as possible. We also propose a normalization technique to compensate for non-uniform rates in recovering pixel values. We conducted simulation studies on Gaussian and diagonal deblurring to evaluate the performance of proposed techniques. Simulation results showed that proposed techniques improved the performance of two-phase methods, by suppressing defective data artifacts effectively.