Robustness Analysis of CNN and Convolutional KAN Architectures against Weather and Geometric Distortions in Traffic Sign Recognition

Abstract

This paper investigates the robustness of traffic sign
classification models against real-world visual disturbances. We
conduct a comparative evaluation of three distinct architectures:
a standard CNN, a hybrid CNN enhanced with Kolmogorov-
Arnold dense layers (CNN-KAN), and a fully convolutional
Kolmogorov-Arnold Network (CKAN). Unlike traditional CNNs,
the KA-based models utilize learnable activation functions, potentially
offering improved resilience. The experiments were conducted
using the German Traffic Sign Recognition Benchmark
(GTSRB) dataset, containing 43 classes of traffic signs. Models
were trained and tested on both original images and versions
degraded by controlled disturbances, including rotation, blur,
brightness variation, and simulated rain. The results demonstrate
that the proposed CNN-KAN model provides consistently superior
performance under small-to-moderate rotations (up to 20
degrees) and moderate brightness increases, achieving the highest
accuracy in all rain-mask scenarios. It remains competitive
under blur, where it ranks second only to the standard CNN.
Performance decreases were observed only at extreme brightness
levels, where both the standard CNN and CKAN maintained
higher stability. Overall, the findings highlight the potential of
Kolmogorov-Arnold-based architectures for improving robustness
in traffic sign recognition systems operating under realistic
and dynamically changing environmental conditions.