Abstract

Assuming with high-quality visual inputs, past research has suggested that gaze allocation in free-viewing of natural scenes is determined by both bottom-up local image saliency computation and top-down semantic and global scene statistics factors. However, given perceptual processing strategy changes with the external noise level and idealistic high-quality images could under-estimate top-down predictive modulation in interpreting retinal inputs, we need to revisit this framework with realistic visual signals often embedded in natural distortions (e.g. rain, snow, fog). In this study, we applied Gaussian low-pass filter, circular averaging filter and Additive Gaussian white noise to systematically distort both man-made and natural landscape scenes, and recorded participants’ gaze patterns in evaluating the perceived quality of the distorted images. Our analysis showed that in comparison with original high-quality images, distorted images with decreasing quality would gradually attract less number of fixations but longer fixation durations, shorter saccades and less scatter of landing positions around the image centre. Interestingly, this systematically varied gaze pattern in scene viewing was determined by the perceived image quality rather than distortion methods, and the same distortion type could have different impact on the perceived image quality and associated gaze pattern to different scene categories. These differences are also observed when looking at basic image features that are computed in early stages of human visual processing. We are currently conducting further computational image analysis to determine relative contribution of bottom-up and top-down factors in guiding gaze behaviour in evaluating of degraded natural scenes.