Abstract
Understanding our sensitivity to noise in naturalistic conditions will provide key clues as to the visual system's adaptations to the real world. Determining which characteristics of image noise (kind, amount, or distribution) appear particularly noisy to humans is a key first step. To examine this, we presented pairs of images (natural landscapes: m=102 shot at ground level; m=289 overhead), which were noise-added versions of the same noise-free image and had observers (n=108 total) judge, in a 2AFC paradigm, which image appeared clearer (less noisy). Varying level and kind of noise, data from 2346 image pairs were obtained. Classifiers were trained to simulate human choice data based on the values extracted for each of the image's forty-four pre-determined features. Features are local or global; local meaning over a small neighborhood around each pixel; global meaning ensemble values over the entire image. Results shown here are the averages of four algorithms (decision tree, random forest, logistic regression, SVM). Ten-fold cross-validation was used. Custom code for feature extraction was written in MATLAB; supervised machine learning was implemented on Pycharm. Our reasoning was to use machine learning to mimic human selection, then leverage the classifier to find what image features could possibly underlie human sensitivity to noise. While the all-features classifier matched (95.3±1.2%) human choice data, its performance was nearly matched by a classifier (93.0±1.3%) based on a single local feature, LocalContrastNearestNeighbor-Red – the local (red) contrast between a pixel and its nearest neighbors. The second-best (92.4±1.4%) single-feature classifier was also local. On the other hand, the best single global-feature classifiers, namely NumberOfOutlines (88.0±1.6%), HistogramMean-Red (83.3±1.9%) and HistogramSkewness-Red (82.8±1.9%) did not perform as well. Generally speaking, clearer images as judged by human observers were of lower local contrast and higher positive skewness than their noisier counterparts. Additional signatures of human-like performance are being explored.