Generate Realistic Traffic Sign Image using Deep Convolutional Generative Adversarial Networks

Abstract

Convolutional Neural Networks (CNNs) achieve excellent in traffic sign detection and recognition with sufficient annotated training data. The quality of the whole vision system based on neural networks depend on the dataset. However, it is complicated to find traffic sign datasets from most of the countries of the world. In this context, Deep Convolutional Generative Adversarial Networks (DCGAN) can synthesize realistic and diverse additional training images to fill the data lack in the real image distribution. This paper mainly discusses how is the quality of the pictures generated by the DCGAN with various parameters. We use an image with a different number and size for training. Further, the Structural Similarity Index (SSIM) and MSE were used to evaluate the quality of the image. Our work measured SSIM values between generated images and corresponding real images. The generated images show high similarity with the actual image while using more images for training. The highest SSIM values reached when using 200 total images as input and images size 32×32.