基于多尺度特征融合的轻量级目标检测算法  

Lightweight target detection algorithm based on multi-scale feature fusion

作　　者：李校林[1,2] 陈泽 LI Xiaolin;CHEN Ze(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;Research Center of Digital Intelligence Technology Application,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065 [2]重庆邮电大学数智技术应用研究中心,重庆400065

出　　处：《微电子学与计算机》2024年第9期32-40,共9页Microelectronics & Computer

摘　　要：由于YOLOv5目标检测模型中参数多、计算复杂度高,无法满足边缘设备进行智能计算和实时反馈的需求,提出了基于多尺度特征融合的轻量级目标检测算法。首先,针对标准卷积模块参数量大、计算复杂度高的问题,提出基于幻影卷积的特征提取卷积模块代替原模型的特征提取模块,在保持检测精度的前提下,减少模型的参数量和计算量。其次,设计出ShuffleNetv2_2下采样模块,进一步减少算法的参数量。再次,针对模型轻量化后的特征提取能力不足问题,将低维特征充分融合到Neck网络中并添加跨层级联以降低浅层语义的丢失,在增强目标特征的表达的同时提高模型的检测效率。最后,提出LAM注意力融合模块,为模型的颈部网络提供具有更丰富的语义特征图。实验结果表明,相比于原模型,改进模型的参数量和计算量更少,并且在PascalVOC和MSCOCO数据集的检测准确率分别提高了2.1%和2.4%。In order to solve the problem of many parameters and high computational complexity in the YOLOv5 target detection model,which cannot meet the needs of edge devices for intelligent computation and real-time feedback.A lightweight target detection algorithm based on multi-scale feature fusion is proposed.Firstly,to address the problem of large number of parameters and high computational complexity of the standard convolution module,a phantom convolutionbased feature extraction convolution module is proposed to replace the feature extraction module of the original model,which reduces the number of parameters and the computational amount of the model under the premise of maintaining the detection accuracy.Then the ShuffleNetv2_2 downsampling module is designed to further reduce the number of parameters of the algorithm.Secondly,to address the problem of insufficient feature extraction ability after model lightweighting,the low-dimensional features are fully fused into the Neck network and cross-layer cascade is added to reduce the loss of shallow semantics,which enhances the expression of the target features and improves the detection efficiency of the model at the same time.Finally,the LAM attention fusion module is proposed to provide the model's Neck network with a richer semantic feature map.The experimental results show that the improved model has fewer parameters and less computation than the original model,and the detection accuracy is improved by 2.1%and 2.4%in the Pascal VOC and MS COCO datasets,respectively.

关 键 词：目标检测 轻量化级神经网络 注意力机制 多尺度特征融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]