Robust Blind Text Image Deblurring via Maximum Consensus Framework

Blur in photographed or scanned text documents is a frequent and damaging degradation that reduces character legibility and OCR accuracy. This paper addresses blind text image deblurring, where neither the blur kernel nor its parameters are known in advance. The method builds a Text-Specific Hybrid Dictionary (THD) from three categories of patch pairs-Gaussian blur-sharp, motion blur-sharp, and sharpsharp-that together cover the range of states an iterative solver encounters as it refines its latent image estimate. A Text Property (TP) Enhancement Operator, based on anchored regression on the trained dictionary, projects each intermediate patch toward the space of sharp text at every iteration. The overall objective is minimized by alternating between latent image and blur kernel subproblems using half-quadratic splitting. Final restoration uses a non-blind deconvolution step once the kernel is reliably estimated. Experiments on both synthetic and real-world blurred text images show consistent improvements over existing methods in PSNR, SSIM, and kernel similarity, with a corresponding gain in OCR accuracy. The experimental results demonstrate that the proposed approach effectively restores blurred text images and significantly improves text readability and OCR performance compared with existing techniques