Abstract
We studied the performance of V1-based models (incorporating both linear and non-linear characteristics) in predicting how observers perceive changes in natural scenes. Pilot studies found that the simplest model was able to predict subjective perception of many types of suprathreshold changes, but consistently underpredicted the actual perceived magnitude of changes in blur. Blur might be perceived differently from other types of changes because it serves as an important cue for accommodation, depth, and motion. This study investigated whether the poor predictions for blur changes arise from differences in higher level processing. To investigate the role of high-level processing in blur perception, we compared our low-level model's performance for blur in eight normal (N) natural scenes and their distorted (D) counterparts (in which the higher-level cues were removed by blurring only selected portions of the natural scenes). If blur change perception were independent of high-level processing, human and model performance should be similar in both conditions. Blur detection thresholds were collected from 3 observers using a 2AFC protocol. Suprathreshold discrimination measurements were obtained using a matching procedure: the blurriness of a test pair was adjusted to match the degree of change in a “comparison pair”. Three-way ANOVAs showed that blur detection and suprathreshold perception were independent of the stimulus type (N versus D): F(2,10)=1.42, P=0.287 and F(1,135)=1.42, P=0.235 respectively. This suggests that higher-level cues did not significantly influence the perception of blur differences. A successful model should give identical outputs for all threshold-level differences. Models of low-level processes in V1 failed to explain the observers' high sensitivities to blur changes (Three-way ANOVA: F(2,10)=6.21, P=0.0177). However, additional modelling of attention and bias towards high spatial frequencies produced some significant improvements (Three-way ANOVA: F(2,10)=1.41, P=0.288). These results suggest that purely low-level models cannot readily describe blur perception, and must incorporate more complex mechanisms.