UC Santa Cruz

While digital imaging systems have been widely used for many applications including consumer photography, microscopy, aerial photography, astronomical imaging, etc., their output images/videos often suffer from spatially varying blur caused by lens, transmission medium, post processing algorithms, and camera/object motion. Measuring the amount of blur globally and locally is an important issue. It can help us in removing the spatially varying blur, and enhancing the visual quality of the imaging system outputs. It can also provide useful information about the scene, such as saliency and depth map.

In this work, we study the blur measurement problem for different scenarios. We first analyze the behavior of a local gradient-based sharpness metric in the presence of spatially varying blur and spatially constant noise, and develop two multiframe restoration systems based on this metric. The first one is a focus-stacking method developed for multifocus image sequences. It selects sharp local patches from the input sequence through the metric, and fuses them together to generate a high quality image. Different from existing stacking methods, this one utilizes image color information to correct fusion errors caused by inaccurate sharpness measurement. The second system aims to restore videos distorted by air turbulence. Air turbulence produces random blur and geometric deformation. Our system first reduces the space and time varying deblurring problem to a shift invariant one through a registration and fusion procedure, and then removes the blur using a blind deconvolution step. Experiments illustrate that this approach can effectively alleviate turbulence distortions, and recover details of the scene.

We then consider the situation where both blur and noise are spatially changing. A quality metric based on singular value decomposition of local image gradient matrix is proposed. Compared with existing sharpness metrics that cannot distinguish quality decay against noise, the proposed one is properly correlated with the noise level and blurriness of a given image. Ample simulated and real data experiments illustrate that this metric is capable of capturing the trend of quality change during the denoising process, and can be used to automatically select the denoising filter parameters that show excellent visual performance in balancing between noise suppression and detail preservation.

Finally, we propose a method capable of estimating the accurate amount of local defocus blur from a single image. This method measures the probability of local defocus level. It also takes smoothness and color edge information into consideration to generate a coherent map indicating the amount of blur at each pixel. Real data experiments illustrate its good performance, and its successful applications in foreground/background segmentation.