In blind image deblurring task, imperfect blur kernel estimation causes severe artifacts in deblurring results. One of the main reasons for the deblurring artifacts is an incorrect selection of the blur kernel size. In this paper, we propose a method for estimating the size of the blur kernel from the given blur image. In this method, we learn the Histogram of Oriented Gradients (HOG) features for sharp images from the corresponding blurry images. For this purpose, we train SVR models using a set of sharp and blurry pair images. This process is performed in different image pyramid levels. In contrast to previous methods, we train a regression model to estimate directly the kernel size from HOG features, without making any assumption on the shape of kernel blur. As our experimental results show, this method better estimates the size of blur kernel from blurry images. This reduces the deblurring artifacts caused by incorrect estimating the kernel size.