Traffic Sign Recognition with Convolutional Kolmogorov-Arnold Networks
Szymon Sołtysiak, Paweł Biernacki, Urszula Libal · 2025
This paper investigates traffic sign classification, a well-established problem within the field of image recognition. Traffic sign classification plays a crucial role in advanced driverassistance systems, traffic monitoring systems, and autonomous vehicles, which are rapidly gaining prominence in modern transportation. However, accurately classifying traffic signs presents significant challenges due to the inherent variability in sign designs, varying lighting conditions, and the presence of adverse weather conditions. This study explores the application of convolutional neural networks (CNNs) for traffic sign recognition, comparing three distinct architectures: 1) a standard CNN, 2) a CNN with Kolmogorov-Arnold dense layers (CNNKAN), and 3) a convolutional network incorporating Kolmogorov-Arnold layers within the convolutional stage (CKAN). CNNs excel in image analysis due to their ability to effectively extract relevant features. Unlike CNNs where activation functions (like ReLU or sigmoid) are fixed, KA networks offer the advantage of learnable activation functions on the connections between neurons. The experiments were conducted on the German Traffic Sign Recognition Benchmark (GTSRB) dataset, comprising 43 classes of German traffic signs with$32 \times 32$pixel images. To evaluate the robustness of each architecture, we employed various image distortion techniques, including rotations, blurring, brightness adjustments, and simulated weather effects such as rain. These interferences were introduced to assess the performance of each model under challenging real-world conditions.