Deblur

Deblur, in the context of imaging, refers to the process of recovering a sharp image from a blurred observation. Blur arises from factors such as camera motion, defocus, atmospheric turbulence, or lens imperfections, and is commonly modeled as a convolution of a latent image with a blur kernel or point spread function, plus noise.

Deblurring is an inverse problem that is typically ill-posed: multiple sharp images can produce the same blurred image, and noise can be amplified in the recovery. Two main problem settings are recognized. Non-blind deconvolution assumes the blur kernel is known; blind deconvolution attempts to estimate both the sharp image and the kernel.

Classical approaches include inverse filtering and Wiener filtering, which are linear methods sensitive to noise. Iterative methods such as Richardson–Lucy deconvolution, often combined with priors, and regularization techniques like total variation, are widely used to stabilize solutions.

More recently, data-driven methods have become dominant. Deep learning models trained on paired blurred and sharp data can produce high-quality results, often generalizing across different blur types. Notable directions include end-to-end deblurring networks and approaches that jointly estimate the blur and the sharp image.

Applications span consumer photography, surveillance, astronomy, and biomedical imaging, wherever motion or defocus blur degrades image quality.

Evaluation commonly uses metrics such as PSNR and SSIM, alongside qualitative perceptual assessment. Deblurring can introduce artifacts, alter textures, or amplify noise, especially when the blur kernel is misestimated or the scene is highly noisy.

Benchmark datasets include synthetic and real-blur collections, such as the GoPro deblurring dataset and earlier blind deconvolution datasets. Despite progress, deblurring remains challenging for complex, spatially varying, or heavily noisy conditions.