Abstract: A no-reference and training-free algorithm was proposed to investigate the detection and segmentation of partial blur for natural images.First,the test image was re-blurred by a Gaussian low-pass filter.Then pixel-wise discrete cosine transformations within micro neighborhoods for both the test image and the re-blurred image were conducted to obtain the blurriness distribution map.Finally,combined with the K-Means clustering algorithm and the morphologic closing operation,the test image could be segmented into blur region and non-blur region.A series of natural images containing out-of-focus,object motion and different complexity were examined.Experiment results demonstrate that the proposed approach can effectively detect and segment the partial blur image and behave well at precision,recall and F-score.Moreover,the proposed approach has a relatively strong consistency with human judgment and a relatively high time efficiency due to no data training.