We propose a new traffic-sign detection method based on a weakly-supervised multi-purpose single convolutional neural architecture. The base classification network used is the very light convolutional architecture, MobileNetv2, which is used as a region proposal network, in addition to being a classification network. The method is divided into two stages; in the first stage MobileNetv2 is trained to suggest certain regions within the image to classify, while in the second stage it is trained to be a traffic-sign classifier. The method attained few milliseconds of processing time for single image or frame testing and averaged about 55 milliseconds on 800x1300 resolution, while maintaining an acceptable accuracy. This method takes advantage of weak supervision which completely eliminates the time required for dataset annotation. We trained and tested our proposed technique on two datasets that have been broadly used for traffic-sign recognition and detection: the German Traffic Signs Recognition Benchmark (GTSRB) and the German Traffic Signs Detection Benchmark (GTSDB).