How many pixels make an image?

Abstract

The human visual system is remarkably tolerant to degradation in image resolution: human performance in scene categorization remains high no matter whether low resolution images or multi-mega pixel images are used. This observation raises the question of how many pixels are required to form a meaningful representation of an image and identify the objects it contains. In this paper, we show that very small thumbnail images at the spatial resolution of 32x32 color pixels provide enough information to identify the semantic category of real world scenes. Most strikingly, this low resolution permits observers to report, with 80% accuracy, 4-5 of the objects that the scene contains, despite the fact that some of these objects are unrecognizable in isolation. The robustness of the information available at very low resolution for describing semantic content of natural images could be an important asset to explain the speed and efficiently at which the human brain comprehends the gist of visual scenes.

Perceptual discrimination of computer generated and photographic faces

“…This surprising accuracy is consistent with previous studies which showed that images as small as 32 × 32 pixels provide enough information to identify objects and the semantic category of real-world scenes [12].…”

On the functions, mechanisms, and malfunctions of intracortical contextual modulation

“…In natural images this contextual guidance is usually sufficient to resolve all or most of the ambiguity. In the limit, contextual disambiguation can even support the 'identification' of objects in degraded images when they are effectively represented only by a single pixel (Torralba, 2009). Contextual disambiguation by amplifying modulation increases the salience of signals as their probability given the context increases.…”

“…Many studies of object recognition have shown that there are strong biases toward interpretations that are more probable given the context (Bar, 2004;Torralba, 2009;Oliva and Torralba, 2007). Effects of these contextual biases range from facilitation of object-naming at the subordinate level (Davenport and Potter, 2004) to the rapid detection of super-ordinate object categories such as animals (Fize, Cauchoix and Fabre-Thorpe, 2011).…”

Autonomous exploration using rapid perception of low-resolution image information

“…To this end, it is interesting to note that our results are similar to those of Torralba and colleagues. Their independent work on determining the spatial resolution limit for scene recognition [39,40] has established through psychovisual experiments that 32 × 32 is sufficient for the identification of semantic categories of real world scenes. Our result of 32 × 24 is a close approximation governed by the desire to preserve the aspect ratio of the original image.…”