Abstract: Blur detection is an important yet challenging task in computer vision. The previous algorithms are mostly difficult to achieve a cost-benefit balance and their performance is largely limited when faced with the blur image polluted by noise. To address these issues, a fast and unsupervised blur detection algorithm is proposed, which is robust to noise. First, a re-blur strategy and Walsh transform are utilized to analyze the input image in sequency domain. Meanwhile, the low-sequency zone of sequency spectrum is adaptively truncated to eliminate the noise interference. Then, a noise-robust local blur metric is constructed and pixel-wise blurriness is calculated to obtain the blur map. Finally, the blur region is segmented using the clustering-guided multi-scale growth framework. Experimental results on CUHK and DUT datasets demonstrate that the proposed algorithm can detect the image blur effectively and efficiently, and achieves the state-of-the-art performance on multiple indicators like precision, recall, F1-measure, mean absolute error, and mean runtime. Especially on noise-polluted conditions, the proposed algorithm significantly surpasses other competitive algorithms.