A non-reference blur image quality evaluation algorithm is proposed.Firstly，the reference image is produced through the Gauss low-pass filter transformation，then the singular value decomposition is made for the image before and after gauss low-pass filter transformation，and then by constructing the formula the change of singular value is computed，and it is made as the image feature vector，finally the image feature vector is trained by applying generalized regression neural network model，the blur value of image is gotten.A lot of experimental results in the three image database show that the new method is simple.Prediction scores and subjective scores have good consistency.The algorithm has obtained a good evaluation index.