The traditional dark channel prior has been successfully applied to the single image deblurring problem. However, the dark channel prior is ineffective for the blur kernel estimation, when the blurry image has significant noise. Motivated by the success of the fractional-order calculation which can nonlinearly suppress the noise of the signal and enhance the low-frequency part of the signal, the fractional-order calculation theory is combined with the dark channel prior. And then, this paper proposes a method of motion blur kernel estimation based on an improved dark channel prior. Firstly, combined with the maximum posterior estimation algorithm and the fractional order dark channel prior, the blur kernel estimation model is constructed. Then, the half-quadratic splitting method is used to solve the non-convex problem of the model. Finally, according to the coarse-fine strategy, the multi-scale iterative framework is used to iteratively estimate the blur kernel and the latent image is estimated by the non-blind image deblurring method. Experimental results show that the proposed algorithm can obtain a better blur kernel in the blurry image with or without significant noise, and can reduce image noise and ringing artifacts, and improve the quality of the latent image estimation, although it need more running time for processing one image. Meanwhile, the proposed method can be used to process different types of blurry image.