Traffic sign detection is an important part of autonomous driving, and its recognition accuracy and speed are directly related to road traffic safety. Although convolutional neural networks (CNNs) have made certain breakthroughs in this field, in the face of complex scenes, such as image blur and target occlusion, the traffic sign detection continues to exhibit limited accuracy, accompanied by false positives and missed detections. To address the above problems, a traffic sign detection algorithm, You Only Look Once-based Skip Dynamic Way (YOLO-SDW) based on You Only Look Once version 8 small (YOLOv8s), is proposed. Firstly, a Skip Connection Reconstruction (SCR) module is introduced to efficiently integrate fine-grained feature information and enhance the detection accuracy of the algorithm in complex scenes. Secondly, a C2f module based on Dynamic Snake Convolution (C2f-DySnake) is proposed to dynamically adjust the receptive field information, improve the algorithm’s feature extraction ability for blurred or occluded targets, and reduce the occurrence of false detections and missed detections. Finally, the Wise Powerful IoU v2 (WPIoUv2) loss function is proposed to further improve the detection accuracy of the algorithm. Experimental results show that the average precision mAP@0.5 of YOLO-SDW on the TT100K dataset is 89.2%, and mAP@0.5:0.95 is 68.5%, which is 4% and 3.3% higher than the YOLOv8s baseline, respectively. YOLO-SDW ensures real-time performance while having higher accuracy.

In multi-domain neural machine translation tasks, the disparity in data distribution between domains poses significant challenges in distinguishing domain features and sharing parameters across domains. This paper proposes a Transformer-based multi-domain-aware mixture of experts model. To address the problem of domain feature differentiation, a mixture of experts (MoE) is introduced into attention to enhance the domain perception ability of the model, thereby improving the domain feature differentiation. To address the trade-off between domain feature distinction and cross-domain parameter sharing, we propose a domain-aware mixture of experts (DMoE). A domain-aware gating mechanism is introduced within the MoE module, simultaneously activating all domain experts to effectively blend domain feature distinction and cross-domain parameter sharing. A loss balancing function is then added to dynamically adjust the impact of the loss function on the expert distribution, enabling fine-tuning of the expert activation distribution to achieve a balance between domains. Experimental results on multiple Chinese-to-English and English-to-French datasets demonstrate that our proposed method significantly outperforms baseline models in both BLEU, chrF, and COMET metrics, validating its effectiveness in multi-domain neural machine translation. Further analysis of the probability distribution of expert activations shows that our method achieves remarkable results in both domain differentiation and cross-domain parameter sharing.

Autonomous driving technology has entered a period of rapid development, and traffic sign detection is one of the important tasks. Existing target detection networks are difficult to adapt to scenarios where target sizes are seriously imbalanced, and traffic sign targets are small and have unclear features, which makes detection more difficult. Therefore, we propose a Hybrid Feature Fusion Traffic Sign detection algorithm based on YOLOv7 (HFFT-YOLO). First, a self-attention mechanism is incorporated at the end of the backbone network to calculate feature interactions within scales; Secondly, the cross-scale fusion part of the neck introduces a bottom-up multi-path fusion method. Design reuse paths at the end of the neck, paying particular attention to cross-scale fusion of high-level features. In addition, we found the appropriate channel width through a lot of experiments and reduced the superfluous parameters. In terms of training, a new regression loss CMPDIoU is proposed, which not only considers the problem of loss degradation when the aspect ratio is the same but the width and height are different, but also enables the penalty term to dynamically change at different scales. Finally, our proposed improved method shows excellent results on the TT100K dataset. Compared with the baseline model, without increasing the number of parameters and computational complexity, $A{P}_{0.5}$ and $AP$ increased by 2.2% and 2.7%, respectively, reaching 92.9% and 58.1%.