Abstract
Blur is a fundamental property for image and optical quality assessment. Blur has been studied with single contours, but natural scenes are composed of a range of depth planes giving rise to retinal images with broad distributions of blur. To study blur perception under more natural conditions, we generated locally controllable dead leaves stimuli – 128 mutually occluding ellipses of random luminance, contrast, orientation, size, aspect ratio, and position. Each element was individually Gaussian blurred allowing blur mean and blur variance to be manipulated independently. Four blocked mean blurs (μ = 2, 4, 8, 16 cycles/image) and three blur variance levels (σ = 0, 0.25, and 0.5 * μ) were interleaved in a 2IFC blur discrimination task. In a matching task, the perceived blur of a high variance image, with fixed mean blur, was matched to that of a low variance image of adjustable mean blur. Matching results and equivalent noise analysis on the blur discrimination data showed that observers were surprisingly capable of integrating wide distributions of blur with limited bias toward sharp or highly blurred elements. Thus, the distribution of local image blur, rather than the blur of single items, determines perceived optical and image quality.