In the application of traffic sign recognition, most of the targets to be detected are small targets, which are prone to problems such as missed detection and false detection. In order to solve these problems, an improved traffic sign recognition algorithm, FKDS-YOLOv8s, was designed based on the YOLOv8s algorithm. FasterBlock is used to reconstruct the C2f module to form a new lightweight module C2f-Faster, which not only improves the feature extraction ability of the model, but also reduces the computational overhead. Based on the SENet and ResNeXt models, a new detection head Detect_SR was designed to enable the model to effectively focus on the key features of small targets. DySample, a lightweight and efficient dynamic upsampler, significantly reduces GPU memory consumption. By increasing the output level of upsampling and prediction, the model can capture rich position information, which effectively solves the problem of insufficient information when the YOLOv8s model processes small targets. The Shape-IoU loss function is introduced to optimize the shortcomings of the original CIoU in the border regression. In addition, the newly designed attention mechanism DKN-Attention is integrated into the Neck part, which locates the attention area of the scene of small objects during the upsampling and downsampling process, and improves the feature extraction and recognition ability of small traffic signs in the distance. The experimental results were carried out on the Chinese traffic sign dataset TT100K, and the results showed that FKDS-YOLOv8s improved the accuracy (P)、recall rate (R) and mAP50 by 5.9%、4.2% and 6.3%, respectively, compared with the benchmark model. Compared with the traditional method, FKDS-YOLOv8s shows significant advantages in performance.