Abstract
Face recognition based on deep neural networks has achieved great success, but its application in resource-constrained and unconstrained scenarios, such as vehicle images from traffic monitoring systems, remains challenging. These scenarios involve complex image variations and require effective training models. To address these challenges, we constructed a low-resolution driver face detection and recognition model, named the SPFL-DC framework, tailored for complex environments. Our primary contribution is a novel method that automates dataset construction using triplet loss, guided by license plate information and pre-training on public datasets. This approach enhances the efficiency of dataset construction and the robustness of model predictions. Secondly, our model used a super-resolution technique to process low-resolution images, fusing them with the original low-resolution inputs. This method improved image resolution while minimizing the loss of critical identity information. We conducted experiments on the AR and LFW datasets to demonstrate the effectiveness of our model, showing competitive performance against state-of-the-art methods.