Autoencoder image denoising as a preprocessing method for traffic sign recognition

Sammanfattning

The ability to recognize traffic signs is essential for maintaining road safety. Traffic sign recognition systems, developed utilizing machine learning, assist drivers with detecting and recognizing traffic signs. However, the robustness of traffic sign recognition systems is heavily affected by real-world variability. Therefore, one potential improvement for traffic sign recognition is preprocessing the images passed to the classification model by reducing noise noise. We developed a denoising autoencoder for this purpose and benchmarked it using all combinations of real and denoised training images as well as real and denoised test images of a dataset consisting of German traffic signs. We found a decrease in accuracy in all cases compared to using real training images with real test images which had an accuracy of 95.2%. Among the cases using denoised data, we found that using denoised training images with real test images had the best performance with an accuracy of 94.8%, and the case using real training images with denoised test images having the worst performance with an accuracy of 82.9%.