Abstract
Contour detection is a crucial part of early vision and is thought to underlie many visual functions. A broad class of computational models (Li, 1998; Ross et al., 2000; Ben-Shahar & Zucker, 2004) implements contour integration by enhancing edge configurations in the visual field that are consistent with natural image statistics (Geisler et al., 2001). These configurations are also known as the "association field" in psychophysics (Field et al., 1993) or the "co-circularity condition" (Hunt et al., 2011). This mechanism seems consistent with the lateral excitatory connections found between orientation-tuned cells of the primary visual cortex (Bosking et al., 1997). Within this framework, inhibition is often introduced simply as a regulatory mechanism to prevent runaway activity. We argue that the two main forms of inhibition, i.e., subtractive and divisive, have more specific interpretations in the context of a contour detection task. Specifically, subtractive inhibition may implement a winner-take-all mechanism, which suppresses edges not part of contours and reduces false alarms. Divisive inhibition, also known as gain control or normalization (Carandini & Heeger, 2011), may scale the neural activity locally according to the maximally responding edge detector, preventing the suppression of contours by stronger ones, thus reducing misses. Here we implemented a neural network model of lateral connections and dual inhibition in V1, optimizing parameters for optimal contour detection. We evaluated the model on a dataset of natural curve fragments (Guo & Kimia, 2012), used in computer vision to evaluate the performance of a bottom-up algorithm for contour integration. The model is shown to perform better than state-of-the-art computer vision systems, such as Pb (Martin et al., 2004). Our results suggest that the diversity of cortical inhibition is a key element of early vision, as suggested by recent neurophysiological evidence (Lee et al., 2012), and will help to build better artificial vision systems.
Meeting abstract presented at VSS 2013