摘要
针对智能交通识别系统需要具备较高的检测速度和识别精度的要求,在YOLOv4-tiny算法的基础上提出一种基于多尺度与坐标注意力机制融合的改进型轻量化YOLOv4-3RSCtiny算法。首先将主干网络中的Resblock_body模块改进为参数量更少的ResblockD轻量化模块,用于提高算法的检测速度;其次引入特征金字塔池化网络,丰富深层特征图的空间信息,在预测阶段引入坐标注意力机制,降低背景信息的干扰;最后利用具有多次跨级融合的路径增强特征金字塔网络,提高算法对小型目标物体的识别率。在TT100K数据集上进行测试,实验结果表明,相较于YOLOv4-tiny算法,YOLOv4-3RSCtiny算法具有较高的准确性和较好的实时性。
Aiming at the requirement of high detection speed and recognition accuracy for intelligent traffic recognition system,an improved lightweight YOLOv4-3RSCtiny algorithm based on the fusion of multi-scale and coordinate attention mechanism is proposed on the basis of YOLOv4-tiny algorithm.Firstly,the Resblock_body module in the backbone network is improved into a ResblockD lightweight module with fewer parameters as a way to improve the detection speed of the algorithm.Secondly,the feature pyramid pooling network is introduced to enrich the spatial information of the deep feature maps.Next,the coordinate attention mechanism is introduced in the prediction stage to reduce the interference of the background information.Finally,the path enhancement algorithm is proposed by using the path enhancement algorithm with the path-enhanced feature pyramid network with multiple cross-level fusion to improve the recognition rate of the algorithm for small target objects.Finally,by testing the algorithm on the TT100K dataset,the experimental results show that the YOLOv4-3RSCtiny algorithm has a better performance compared to YOLOv4-tiny,with higher accuracy and better real-time performance.
作者
胡腾
杨毅强
邹显迪
孙潇
毛国斌
HU Teng;YANG Yi-qiang;ZOU Xian-di;SUN Xiao;MAO Guo-bin(School of Automation and Information Engineering,Sichuan University of Science&Engineering,Sichuan Yibin 644000,China;School of Computer Science&Engineering,Sichuan University of Science&Engineering,Sichuan Yibin 644000,China)
出处
《齐齐哈尔大学学报（自然科学版）》
2024年第5期8-15,共8页
Journal of Qiqihar University(Natural Science Edition)
基金
人工智能四川省重点实验室项目(2023RYY06)
四川省科技厅项目(2022ZHCG0035)。
