Abstract: To address the issues in traffic sign detection where distant small object features are easily weakened and difficult to distinguish from complex backgrounds, this paper proposes an improved YOLOv10s-based model named DTI-YOLO. Firstly, the PSA module is replaced with the dilated convolution and dilated attention fusion module (DDFM), designed with local and global feature extraction branches that focus on local details and global semantics, respectively, helping to suppress background noise and enhancing the model’s ability to separate small object features from complex scenes. Secondly, a two detection layer-based cross-scale feature fusion network (TDL-CCFN) is constructed, which integrates the cross-scale fusion structure and the two detection layer mechanism tailored for small objects, improving feature fusion between deep and shallow layers, enhancing small object feature retention, and reducing model parameters. Finally, the InnerMPDIoU loss function replaces the CIoU loss, combining an adjustable scale factor with a vertex-based geometric distance metric to enhance the model’s sensitivity to the position and perspective changes of small objects, thereby improving bounding box regression performance and generalization capability. Experimental results demonstrate that, compared to YOLOv10s, the proposed DTI-YOLO achieves excellent detection performance and improves mAP50 by 5.5 and 4.8 percentage points on the TT100K and CCTSDB datasets respectively, reaching 90.9% and 86.6%, while the number of parameters is reduced by approximately 33.3%, to 5.4×106, achieving significant model lightweighting.

Key words: traffic sign detection, DTI-YOLO, complex background, small object features, multi-scale characteristics

摘要： 针对交通标志检测中，远景小目标特征易被弱化，难以与复杂背景区分的问题，提出了一种基于改进YOLOv10s的交通标志检测模型（DTI-YOLO）。提出膨胀卷积融合膨胀注意力模块（DDFM）替换PSA模块，设计局部和全局特征提取分支，通过聚焦局部细节与全局语义，抑制噪声干扰，增强模型在复杂背景中分离小目标特征的能力。构建基于二检测层的跨尺度特征融合网络（TDL-CCFN），利用跨尺度特征融合结构和针对小目标设计的二检测层结构，增强深浅层特征间的融合和小目标特征的保留，同时减少了模型的参数量。引入InnerMPDIoU损失函数替换CIoU损失函数，通过融合可调节尺度因子和顶点几何距离度量，增强模型对小目标位置和视角变化的敏感性，提升边界框回归效率与模型泛化能力。实验结果表明，DTI-YOLO模型具有良好的检测性能，相较于YOLOv10s，DTI-YOLO在TT100K和CCTSDB数据集上的mAP50分别提升5.5和4.8个百分点，分别达到90.9%和86.6%；同时，参数量减少约33.3%，降至5.4×106，实现了模型轻量化。

关键词: 交通标志检测, DTI-YOLO, 复杂背景, 小目标特征, 多尺度特征