Enhancing CNNs robustness to occlusions with bioinspired filters for border completion

TL;DR

This paper addresses occlusion in visual classification by implementing border completion mechanisms inspired by the visual cortex, formalized via an orientation map $Theta$ and an orientation vector $Z$ to guide contour integration. The authors augment LeNet 5 with four bioinspired filters at the input (BorderNet) that mimic horizontal, vertical, and diagonal receptive fields, alongside a RandomNet with random filters, trained on unoccluded MNIST. Testing on an occluded MNIST dataset with diagonal stripes of width $w$ and spacing $s$ shows BorderNet outperforms Vanilla LeNet and RandomNet, demonstrating robustness not due to parameter count. The work suggests a general strategy for improving CNN robustness under occlusion by incorporating geometry-aware, bioinspired filters, with future work extending to more occlusion types and datasets.

Abstract

We exploit the mathematical modeling of the visual cortex mechanism for border completion to define custom filters for CNNs. We see a consistent improvement in performance, particularly in accuracy, when our modified LeNet 5 is tested with occluded MNIST images.

Figures (6)

• Figure 1: The association field.
• Figure 2: Oriented (top row) and random (bottom row) filters.
• Figure 3: A sample of occluded images used in the testing phase.
• Figure 4: Accuracies with occlusion stripe spacing and width for the Vanilla LeNet and BorderNet
• Figure 5: Accuracy improvement of BorderNet with respect to Vanilla LeNet.

Theorems & Definitions (1)

• definition thmcounterdefinition