针对复杂场景下目标检测与识别精度较低的问题,提出了一种基于注意力与多级特征融合的 YOLOv5 目标检测与识别算法.该算法在传统YOLOv5s模型的主干网络中引入双空间方向的金字塔切分注意力机制,增强对特征空间和通道信息的学习能力,同时在瓶颈网络中采用多级特征融合结构,对不同分支的特征进行融合,增加特征的丰富性,提升应对复杂场景的能力.此外,利用 C3Ghost 模块和深度可分离卷积分别替换 C3 模块和普通卷积,降低网络参数量和复杂度.结果表明:与传统的 YOLOv5s算法相比,所提算法在 VOC2007+2012 数据集的均值平均精度高达 85%,在智能零售柜商品识别数据集的均值平均精度高达 97.2%,表现出较好的性能.
To tackle the problem of low accuracy of detection and recognition for object in complex scenes,YOLOv5 object detection and recognition algorithm based on attention and multistage feature fusion(AMFF)was proposed in this study.The main ideas included adding the proposed dual space directions pyramid split attention(DSD-PSA)mechanism to the backbone network of the traditional YOLOv5s model to enhance the learning of the feature map space and channel information,adopting multistage feature fusion(MFF)structure in the bottleneck network to fuse the features of different branches,increasing richness of the feature and improving the ability to cope with complex scenes.In addition,C3Ghost module and depthwise separable convolution were used to replace C3 module and common convolution to reduce the number of parameters and the complexity of network.Compared with the tradi-tional YOLOv5s algorithm,the mean average accuracy of the proposed algorithm in the VOC2007+2012 data set reached 85%,and the mean average accuracy of the smart retail cabinet commodity identification data set reached 97.2%,which verified the effectiveness and feasibility of the proposed algorithm.
王瑜;毕玉;石健彤;肖洪兵;孙梅
北京工商大学 计算机与人工智能学院,北京 100048
计算机与自动化
深度学习;YOLOv5s;目标检测;多级特征融合;注意力机制
deep learning;YOLOv5s;object detection;multistage feature fusion;attention mechanism
《郑州大学学报（工学版）》 2024 (003)
38-45,95 / 9
北京市教委-市自然科学基金联合资助项目(KZ202110011015)
10.13705/j.issn.1671-6833.2023.06.009
评论