Image Blur Classification and Parameter Identification Using Two-stage Deep Belief Networks

Image blur kernel classification and parameter estimation are critical for blind image deblurring. Current dominant approaches use handcrafted blur features [5, 6] that are optimized for a certain type of blur, which is not applicable in real blind deconvolution where the Point Spread Function (PSF) of the blur is unknown. Inspired by the successful applications of deep learning techniques to object recognition and image processing [2, 4], in this paper, a Two-stage system using Deep Belief Networks (TDBN) is proposed to first classify the blur type and then identify its parameters. In this paper, we intend to design a patch-based blur type classification and parameters identification method to better solve the realistic blur analysis problem. Deep Belief Network (DBN) [3] is chosen for accomplishing the feature extraction and final classification in this system. A two-stage framework is proposed: first, for the input image patches with different blurs, the DBN is used for identifying the blur type; second, different samples with the same blur type will be sent to the corresponding DBN blocks for further parameter estimation. The DBN is trained in a semi-supervised way: the unsupervised training of the DBN is done by a greedy layer-wise pre-training before the supervised backpropagation for the fine-tuning. The unsupervised process helps the feature learning, and the backpropagation helps to construct the discriminative information. In a word, our contributions are threefold: 1) To the best of our knowledge, this is the first time that deep belief network has been applied to the problem of blur analysis; 2) A discriminative feature, derived from edge extraction on Fourier Transform coefficients, has been proposed to preprocess blurred images before they are fed into the DBN; 3) A two-stage framework is proposed to estimate the blur type and parameters for any given image degraded by spatially invariant blur of an unknown type. To begin with, for the input blurred patches, the logarithmic spectra are used as a type of feature for the identification of the blur pattern, which is shown in the following Fig. 2. Three types of blur have been considered in our paper, which are the Gaussian blur, the motion blur and the defocus blur. They can be formulated as: