The invention discloses an image motion blur blind removing method based on gradient domain and depth learning, A guide filter gradient domain image is used as a basic image, A L0 filt gradient domainimage and a corresponding clear image are taken as samples, that clear image is randomly convolution with different blur kernel, and 1% Gaussian white noise is added to generate a motion blur image.The gradient domain image after the guide filter, the gradient domain image after the L0 filter and the motion blur image constitute a training data set; The depth convolution neural network is constructed, and the weight data of the depth convolution neural network is learned from the training data set, and the depth convolution neural network used for motion fuzzy kernel estimation is learned. Extracting the weight data of the network training, obtaining the motion blur kernel, optimizing the deconvolution function of the image prior constraint, and obtaining the motion blur image by using the total variational term. This method can effectively suppress the image ringing effect and reduce the image noise, and the effect of motion blur removal is better.