We develop a method to optimize a machine-based semantic categorization of natural images according to human perception. First, the categories are determined through a psychophysical experiment. The similarity matrices obtained from the human responses are analyzed by a multidimensional scaling technique called curvilinear component analysis (CCA). The same is done with an automatic image indexing system based on similarities between the outputs of Gabor filters applied to the images. Then we show that, by using the human categorization to balance the filter outputs, the system's performance may be significantly improved.