An Effective Image Synthesis Method to Solve the Shortage of Traffic Sign Training Images in Taiwan

With the advancement of smart transportation, the integration of traffic sign detection into autonomous driving systems based on machine learning has become increasingly crucial. However, Taiwan has not yet established the corresponding training dataset, making it extremely challenging to acquire sufficient images and corresponding annotations. Additionally, the varying frequencies of different traffic signs result in an imbalance in the quantity of each classes within the dataset. To address these issues, this paper proposes an image synthesis method based on data augmentation techniques, leveraging the similarity between traffic signs and their mockups. This approach involves placing mockups of traffic signs onto various backgrounds to simulate diverse images. Through such synthesized images, the model can be successfully trained and achieved results comparable to those solely trained from real images. This underscores the potential and effectiveness of synthesized images in compensating for insufficient real data, establishing high-precision traffic sign detection models, and creating reliable automatic labeling systems.