Radio frequency fingerprinting is a key technology that plays an important role in enhancing the security of Internet-of-Things applications. In this paper, we present a new radio frequency fingerprinting system based on a novel feature extraction technique. We first convert the collected steady-state signals to grayscale images by byte without the need for any prior knowledge. Next, the collected data is fed into a lightweight neural network called MobileNet for training and classification. To evaluate the performance of the proposed system, we then conduct experiments with 10 Long Range (LoRa) devices and a general software radio receiver. Experimental results show that the proposed model outperforms some mainstream models. Moreover, we input mobile phone device data into our system. Experimental results demonstrate that our proposed model can achieve a significant classification accuracy of 99.23%.