Abstract
Stimulus complexity is relevant for a variety of perceptual phenomena; for example, more complex shapes are harder to detect in a cluttered environment. Lacking a way to concretely measure complexity, such studies frequently rely on subjective ratings of complexity by humans. Here we propose an objective method for measuring the complexity of an image using Multiscale Entropy (MSE). MSE has previously been used to measure the complexity of time series (Costa, Goldberger, & Peng, 2002), measuring regularity across different frequencies. Here we adapt this method for use with images. Given an input image, we compute a Gaussian pyramid, where each layer corresponds to a different spatial scale. We measure the Shannon entropy of the luminance distribution at each scale and combine these scores into a single complexity rating. We isolate the effect of color by computing MSE separately for the hue, saturation, and brightness of each image. We obtained human ratings of complexity for 690 images of real-world nature scenes – approximately 50 ratings per image – using Amazon Mechanical Turk, and compared our method to human ratings. Our method is highly predictive of human complexity ratings for real-world scenes (r = 0.66), outperforming Shannon entropy computed only for the original image (r = 0.38). MSE is simple to implement and has potential applications in many different areas where image complexity is of importance. One such area is aesthetics of images, where the emotional valence of an image has been shown to be related to the complexity of the image. Having a concrete measure of complexity that matches human ratings will allow for a more in-depth study of this relationship, allowing, for instance, for balancing of stimulus sets for visual complexity.