Predicting perturbation of object contours by background in natural images

Abstract

Purpose: Because object contours are composed of local contrast elements, perturbation of the local contrast elements will result in perturbation of an object's contour. Local contrast is generally measured using contrast sensitive filters. Contrast filters can measure both preferred orientation and contrast magnitude. Contrast filters, applied along object boundaries, take input from both the object image and the background image. Changes in the background, thus, cause changes in the filter response. This perturbation can be measured and described statistically. The goal of this study is to predict the structure of contour perturbation statistics, using background summary statistics and develop the corresponding theory. Methods: Images of natural and artificial objects were photographed and hand-segmented from their natural backgrounds. Composite images were made by digitally superimposing and translating the image of an object across various natural image backgrounds. Perturbation statistics were represented using 2D histograms of orientation and magnitude responses. Histograms were organized into a taxonomy, according to their prominent features. Features included single or double histogram modes. Background summary statistics were studied in order to determine if they could help predict the locations of the modes within the histogram. Results: It was found that the mode of the histogram of the background pixel intensities could predict the position of the 2D histogram modes. Prediction of mode location was highly accurate. It was also possible to predict the number of modes. Conclusions: It is possible to predict important features of contour perturbation, using summary statistics of the background.