Impact of Image Resizing Algorithms on Deep Learning Models for Traffic Sign Recognition
Yu-Hsu Lee, Weng Kin Lai, Win Kent Ong · 2025
The ability to accurately identify and decipher traffic signs is a major obstacle on the road to widespread adoption of autonomous vehicles. While advanced algorithms are constantly evolving, the impact of image resizing, a commonly overlooked preprocessing step, on model performance remains underexplored. This research bridges this crucial gap by investigating the influence of six resizing algorithms i.e., nearest neighbour, bilinear, bicubic, Lanczos, box, and Hamming) on five popular neural network architectures i.e., EfficientNet-B0, ResNet-18, Lenet-5, MobileNet and Inception-V3. Evaluations based on the widely used German Traffic Sign Recognition Benchmark (GTSRB) and the Traffic Sign dataset (TS), revealed significant variations in performance based solely on the chosen image resizing technique. These findings not only illuminate a previously unexplored aspect of traffic sign recognition but also offer valuable insights for optimizing model accuracy in self-driving vehicles.