Model Cortical Association Fields Account for the Time Course and Dependence on Target Complexity of Human Contour Perception

Abstract

Can lateral connectivity in the primary visual cortex account for the time dependence and intrinsic task difficulty of human contour detection? To answer this question, we created a synthetic image set that prevents sole reliance on either low-level visual features or high-level context for the detection of target objects. Rendered images consist of smoothly varying, globally aligned contour fragments (amoebas) distributed among groups of randomly rotated fragments (clutter). The time course and accuracy of amoeba detection by humans was measured using a two-alternative forced choice protocol with self-reported confidence and variable image presentation time (20-200 ms), followed by an image mask optimized so as to interrupt visual processing. Measured psychometric functions were well fit by sigmoidal functions with exponential time constants of 30-91 ms, depending on amoeba complexity. Key aspects of the psychophysical experiments were accounted for by a computational network model, in which simulated responses across retinotopic arrays of orientation-selective elements were modulated by cortical association fields, represented as multiplicative kernels computed from the differences in pairwise edge statistics between target and distractor images. Comparing the experimental and the computational results suggests that each iteration of the lateral interactions takes at least ms of cortical processing time. Our results provide evidence that cortical association fields between orientation selective elements in early visual areas can account for important temporal and task-dependent aspects of the psychometric curves characterizing human contour perception, with the remaining discrepancies postulated to arise from the influence of higher cortical areas.

Director Field Model of the Primary Visual Cortex for Contour Detection

“…The model performs on par or better than agent-based three-dimensional models (two spatial dimensions and one orientation preference dimension), with complex, empirically specified co-circular interaction kernel [20]. This illustrates that discreteness of neurons, existence of the orientation preference as an independent variable, and intricate details of the kernel are not crucial for the studied visual processing function.…”

“…The model fills in the occlusions and filters out the clutter based on the presence or absence of co-circular contextual edge support. In addition to the substantial simplification, this ability to fill in the occlusions particularly distinguishes our approach from the previous work on co-circular excitatory feedback [19,20]. It remains to be seen to which extent the performance is affected by more natural statistics of images, and by the presence of stochasticity and synaptic plasticity in neural dynamics.…”

“…Note again the reduced complexity of this model, where the kernel is defined by just two real-valued parameters, instead of being inferred empirically from the data in a form of a multi-dimensional matrix, as in Ref. [20] and references therein. We define the input nonlinearity using a complex step function:…”

“…We do not aim at the state of the art performance on complex natural images, but rather ask what is the smallest set of computational primitives that must be implemented in a model to achieve such detection and integration of long contours in a nontrivial setting. For this, we focus on a proposal of a specific lateral connectivity among V1 neurons [19,20], which incorporates the Hebbian constraint that neurons that are excited simultaneously by the same long, low-curvature contours should activate each other [19]. However, in our model, we do not reproduce the complexity of V1, which has ∼ 100 million neurons, with each neuron having 10 3 connections, some extending for many millimeters.…”

Combining multiple visual processing streams for locating and classifying objects in video

“…Rather than learning a dictionary of features whose complexity increases as one travels up a cortical hierarchy, our implementation of lateral interactions uses a very simple set of feature detectors, corresponding to edge detectors spanning eight orientations between 0 and 180 degrees. The activation of these feature detectors was modulated by extensive lateral interactions based on co-occurrence of edges [16].…”