<![CDATA[Traffic sign recognition (TSR) is critical for systems that rely on its output to inform downstream decisions. This study investigates TSR using Capsule Network (CapsNet), an advancement over the convolutional neural network (CNN) that captures spatial relationships between image features, conferring robustness to affine transformations. The proposed architecture comprises a convolutional layer (256 filters, 9X9 kernel, stride 1, ReLU activation), a primary capsule layer (32 channels of 6X6 capsules, each an 8-dimensional vector), and a class capsule layer (one 16-dimensional capsule per target class). The model was evaluated on both original and augmented datasets, the latter incorporating rotations ranging from -5 degrees to +5 degrees. On the original dataset, CapsNet achieved 100% training and testing accuracy with a training loss of 0,0048 at epoch 20. On the augmented dataset, the model achieved 100% training accuracy (loss: 0,0056) and 98% testing accuracy (5 misclassifications). Compared to multi-scale CNN (MS-CNN), which produced 7 misclassifications on the augmented dataset, CapsNet demonstrated superior consistency and robustness under affine transformations. These findings suggest that CapsNet is a viable and effective architecture for real-world TSR applications.]]>
Copyrights © 2026
