Effects of Blur and Noise on Digital Imagery Interpretability

Abstract

Ten medium scale (1:2700 to 1:4400) aerial photographs typical of the imagery viewed by Air Force photointerpreters, were digitized to 4096 x 4096 files (20-Mm aperture by 11 bits intensity) on a scanning microdensitometer. The image files were then multiplied by two Gaussian filter functions (Fourier domain) to yield two blurred and one “ground truth” level of each image and transformed to eight bits of intensity for output. One of four weightings of a 4096 x 4096 Gaussian noise file was added to each image file, yielding 3 Blur x 5 Noise x 10 Image combinations (150 images, total). Positive transparencies then served as the data base for an information extract task. Fifteen photointerpreters (PIs) from the 548th Reconnaisance Technical Group, Hickman AFB, Hawaii, served as subjects. Blur was a between-subjects variable with five PIs at each of three noise levels. Noise was a with in-subject variable (five levels). The nose main effect was significant (p < .01). The Blur main effect and the Blur x Noise interaction were not found to be statistically significant. The data were correlated with scaling values from a separate study using the same images and PIs. A significant correlation was found (r = .898).