Optimal contour integration: When additive algorithms fail

Contour integration is a fundamental computation during image segmentation. Psychophysical evidence shows that contour integration is performed with high precision in widely differing situations. Therefore, the brain requires a reliable algorithm for extracting contours from stimuli. While according to statistics, contour integration is optimal when using a multiplicative algorithm, realistic neural networks employ additive operations. Here we discuss potential drawbacks of additive models. In particular, additive models require a subtle balance of lateral and afferent input for reliable contour detection. Furthermore, they erroneously detect an element belonging to several jittered contours instead of a perfectly aligned and thus more salient contour.