Representing Blurred Image without Deblurring

The effective recognition of patterns from blurred images presents a fundamental difficulty for many practical vision tasks. In the era deep learning, main ideas to cope with this are data augmentation and deblurring. However, both facing issues such as inefficiency, instability, lack explainability. paper, we explore simple but way define invariants images, without Here, designed Fractional Moments under Projection operators (FMP), where blur invariance rotation guaranteed by general theorem Fourier-domain equivariance, respectively. general, proposed FMP not only bears simpler explicit definition, also has useful representation properties including orthogonality, statistical flexibility, well combined blurring rotation. Simulation experiments provided demonstrate our FMP, revealing potential small-scale robust problems.