The Role of Early Recurrence in Improving Visual Representations

This dissertation proposes a computational model of early vision with recurrence, termed as early recurrence. The idea is motivated from the research of the primate vision. Specifically, the proposed model relies on the following four observations. 1) The primate visual system includes two main visual pathways: the dorsal pathway and the ventral pathway; 2) The two pathways respond to different visual features; 3) The neurons of the dorsal pathway conduct visual information faster than that of the neurons of the ventral pathway; 4) There are lower-level feedback connections from the dorsal pathway to the ventral pathway. As such, the primate visual system may implement a recurrent mechanism to improve visual representations of the ventral pathway. Our work starts from a comprehensive review of the literature, based on which a conceptualization of early recurrence is proposed. Early recurrence manifests itself as a form of surround suppression. We propose that early recurrence is capable of refining the ventral processing using results of the dorsal processing. Our work further defines a set of computational components to formalize early recurrence. Although we do not intend to model the true nature of biology, to verify that the proposed computation is biologically consistent, we have applied the model to simulate a neurophysiological experiment of a bar-and-checkerboard and a psychological experiment involving a moving contour illusion. Simulation results indicated that the proposed computation behaviourally reproduces the original observations. The ultimate goal of this work is to investigate whether the proposal is capable of improving computer vision applications. To do this, we have applied the model to a variety of applications, including visual saliency and contour detection. Based on comparisons against the state-of-the-art, we conclude that the proposed model of early recurrence sheds light on a generally applicable yet lightweight approach to boost real-life application performance.