Black, uppercase letters, subtending 6.0' of arc in height, were presented tachistoscopically to 6 subjects. An exposure duration was chosen to keep the subject's identification performance at about 50% correct. On each trial a single letter was presented, and the subject was required to identify the letter by verbal response. The resulting 26 X 26 confusion matrix was based on 3,900 trials (150 trials per letter). Several models of visual processing were used to generate predicted confusions among letter pairs. Models based on template overlap, geometric features, and two-dimensional spatial frequency content (Fourier transforms) were tested. The highest correlation (.70) between actual and predicted confusions was attained by the model based on the Fourier transformed letters filtered by the human contrast sensitivity function. These results demonstrate that the spatial frequency content of visual patterns can provide a valuable metric for predicting their psychological similarity. The results further suggest that spatial frequency models of visual processing are competitive with feature analysis models.