Role of noise in image processing by the human perceptive system M. Piana INFM, Unita ` di Ricerca di Genova, I-16146 Genova, Italy M. Canfora Dipartimento di Fisica, Universita ` di Genova, Genova, Italy M. Riani INFM and Dipartimento di Fisica, Universita ` di Genova, Genova, Italy Received 9 February 2000 Two psychophysics experiments are described, pointing out the significant role played by stochastic reso- nance in recognition of capital stylized noisy letters by the human perceptive apparatus. The first experiment shows that an optimal noise level exists at which the letter is recognized for a minimum threshold contrast. A simple two-parameter model that best fits the experimental data is also discussed. In the second experiment we show that a dramatically increased ability of the visual system in letter recognition occurs in an extremely narrow range of increasing noise. Possible interesting future investigations suggested by these experimental results and based on functional imaging techniques are discussed. PACS numbers: 87.10.+e, 05.20.-y, 89.70.+c I. INTRODUCTION Stochastic resonance SRis a statistical phenomenon as- sociated with nonlinear systems 1; for a large class of such systems, an increase in the noise affecting the input signal may induce an increase in the signal-to-noise ratio SNRin the system’s output. The basic ingredients for SR to show up are a coherent small signal, a form of energetic threshold, and the possibility of varying the amount of noise affecting the signal either by external addition or by some inherent process. In this situation it can be shown 2,3that there exists an optimal level of noise maximizing the response of the system in a resonancelike behavior. Experimental verifi- cations of this effect have been performed in many areas of applied physics and in biology 4,5see, for example, 6,7 for recent and rather complete reviews. The interest of such a phenomenon in the processing of information by the bio- logical neural system is evident at all levels from the lower ‘‘physiological’’ levels to the higher ‘‘cognitive’’ ones. For sensory systems SR effects have been explored, for example, in experiments on single mechanoreceptors from crayfish tails 8, crayfish 9and cricket 10ganglion cells, rat hip- pocampal cells 11, and human muscle spingles 12. At higher levels of complexity in cognitive systems, SR effects have been found in a neural network modeling the phenom- enon of perceptual alternation occurring in the observation of the so-called ambiguous pattern 13, in the human tactile system 14, and in human visual perception 15,16. An interesting behavioral experiment is described in 17, and in 18the role of noise in the auditory system is reported. Both these experiments may link stochastic resonance to evolu- tion, that is, the use of SR can obviously have a survival value and thus was selected for. The results of our experi- ments may also point in the same direction. We are inter- ested in the key question whether and how the human brain exploits noise in order to enhance the quality of external stimuli. This problem has been addressed in 15with a psy- chophysics experiment concerning visual perception of noisy patterns. This experiment shows quantitatively that the hu- man brain is helped by noise in detecting small details in stationary images and that this visual enhancement is satis- factorily modeled by a one-parameter SR curve obtained from level-crossing detector theory 19. In the present paper we discuss two experiments con- cerned with the visual perception of noisy letters. More pre- cisely, in experiment I, characterized by an experimental paradigm similar to the one in 15, we produce images con- taining one letter each that we painted over a uniform back- ground and depressed under a fixed threshold, i.e., pixels with gray level lower than the threshold are painted with the same gray level as the background. Then we affect each letter with noise of different standard deviations, and for each presentation we smoothly increase the contrast between the letter signal and the background until the subject recog- nizes the letter. By plotting the contrast value for which the letter is recognized by the subject versus the value of the standard deviation characterizing the presentation, we can show that an optimal noise level, where the recognition con- trast value is minimum, can be detected. As the main results of this experiment we obtain the following. 1SR occurs when the human perceptive apparatus is asked to recognize rather big stylized noisy capital letters previously depressed under a fixed threshold. 2A quantitative estimate of the optimal noise level can be produced for all the subjects. Knowledge of the corre- sponding contrast threshold is helpful for the realization of the second experiment in this paper. 3The theoretical model describing the detection of small details in the experiment in 15is here able to follow only coarsely the trend of the noise effect on the contrast threshold of the human visual system. However, we can pro- vide a two-parameter modification of that model fitting our data in a more reliable fashion. In a second experiment experiment IIeach subject is PHYSICAL REVIEW E JULY 2000 VOLUME 62, NUMBER 1 PRE 62 1063-651X/2000/621/11046/$15.00 1104 ©2000 The American Physical Society