Synthetic Data generation using DCGAN for improved traffic sign recognition

Abstract

Traffic sign detection and recognition perform a vital function in real-world driver guidance applications, including driver assistance systems. Research into vision-based traffic sign detection (TSD) and traffic sign recognition (TSR) has gained considerable attention in the scientific community, led mainly by three variables: identification, monitoring, and classification. In addition, TSR provides valuable details and alerts for smart cars including advanced driving assistance (ADAS) and cooperative intelligent transport systems (CITS). Our work will generate high-quality synthetic prohibitory sign images using deep convolutional generative adversarial networks (DCGAN). This paper analyzes and discusses CNN models incorporating different backbone architectures and feature extractors, focusing on Resnet 50 and Densenet for object detection. Assessment of the models provides important information, including mean average accuracy (mAP), workspace capacity, detection period, and the amount of billion floating-point operations (BFLOPS). The maximum average accuracy is 92% (Densenet DCGAN), led by 91% (Resnet 50 DCGAN), 88% (Densenet), and 63% (Resnet 50). We find when using the original image and a synthetic image, accuracy increases, while detection time falls. Our findings show that combining original images and synthetic images in the dataset for training can improve intersection over union (IoU) and traffic sign recognition performance.