Blur image classification based on deep learning

Blur type identification is significant for blind image recovery in image processing area. In this paper, an accurate classification system exploiting Convolution Neural Network (CNN) is designed to identify four blur types of images: defocus blur, Gaussian blur, haze blur and motion blur. A supervised learning model of Simplified-Fast-Alexnet (SFA), which is an abbreviated and modified version of Alexnet, is created to map the input images into a higher dimensional feature space, in which the blurs can be classified accurately. With proportional compressing the output number of each convolution layer in Alexnet by the ratio of 0.5 and removing the first two Full Connected layers (FCs) in Alexnet, the SFA successfully simplifies the Alexnet and overcomes the fatal disadvantage of parameter redundancy. Moreover, the batch normalization layers are added into the designated classifier to replace the dropout method, thus it can accelerate the convergence rate of deep network during the training stage by reducing internal covariate shift as well as alleviate the overfitting problem. Experiments demonstrate the remarkable performance of the suggested approach in comparison with the original Alexnet and the state-of-the-art on the frequently-used Berkeley dataset and Pascal VOC 2007 dataset.