Blur Invariants for Image Recognition

Abstract Blur is an image degradation that makes object recognition challenging. Restoration approaches solve this problem via deblurring, deep learning methods rely on the augmentation of training sets. Invariants with respect to blur offer alternative way describing and recognising blurred images without any deblurring data augmentation. In paper, we present original theory invariants. Unlike all previous attempts, new requires no prior knowledge type. The invariants are constructed in Fourier domain by means orthogonal projection operators moment expansion used for efficient stable computation. Applying a general substitution rule, combined spatial transformations easy construct use. Experimental comparison Convolutional Neural Networks shows advantages proposed theory.