How much contrast information is needed for reliable fast image recognition?

The visual system faces the problem of extracting biologically-relevant information from a large "ux of input data. This can be obtained by summarizing complex scenes to extract meaningful features (Barlow, 1959; Marr, 1976) by using image primitives (edges, bars), encoded physiologically by speci#c con#gurations of receptive #elds (Hubel & Wiesel, 1962). This work follows a pattern-#ltering approach, based on the principle of e!cient information coding under real-world limitations (Punzi & Del Viva, VSS-2006). The model, applied to black and white images predicts from very general principles the structure of early visual #lters and identi#es salient features (edges, lines) providing highly compressed “primal sketches” of visual scenes (Del Viva & Punzi VSS-2008). Human subject are able to identify such sketches (2AFC procedure) in rapid identi#cation tasks (10–20 ms), with very high accuracy (up to 90%), comparable to that for fully detailed original images (Del Viva et al., VSS-2010). Here, we extended previous computational and psychophysical experiments to gray-level images to investigate whether this early visual image processing can make use of a larger amount of input information. Results with 4 gray level images (2 bits) provided sketches with a lesseror equal level of compression, and comparable information content to those obtained with 1 bit. Performance in recognizing image sketches did not improve by increasing input information either. Our results provide support to the idea that only a very limited contrast information is used for fast image recognition, and that this is fully explained by our model of e!cient information within constraints.