Visually-salient contour detection using a V1 neural model with horizontal connections

TL;DR

This paper presents a biologically-inspired neural model based on V1 cortical activity and horizontal connections to detect salient contours in images, demonstrating improved performance through facilitation mechanisms and iterative feedback.

Contribution

The paper introduces a convolutional V1 neural model incorporating horizontal connections for contour saliency detection, with mechanisms for facilitation and iterative enhancement.

Findings

Model performs well across different contour types.

Facilitation mechanisms vary for different contours.

Iterative feedback improves saliency detection.

Abstract

A convolution model which accounts for neural activity dynamics in the primary visual cortex is derived and used to detect visually salient contours in images. Image inputs to the model are modulated by long-range horizontal connections, allowing contextual effects in the image to determine visual saliency, i.e. line segments arranged in a closed contour elicit a larger neural response than line segments forming background clutter. The model is tested on 3 types of contour, including a line, a circular closed contour, and a non-circular closed contour. Using a modified association field to describe horizontal connections the model is found to perform well for different parameter values. For each type of contour a different facilitation mechanism is found. Operating as a feed-forward network, the model assigns saliency by increasing the neural activity of line segments facilitated by the horizontal connections. Alternatively, operating as a feedback network, the model can achieve further improvement over several iterations through cooperative interactions. This model has no dynamical stability issues, and is suitable for use in biologically-inspired neural networks.