基于通道注意力与空间金字塔的改进型YOLOv3及其应用  

Channel attention and spatial pyramid-based improved YOLOv3 and its application

作　　者：游双 张著洪 YOU Shuang;ZHANG Zhuhong(College of Big Data and Information Engineering,Guizhou University,Guiyang 550025,China;Guizhou Provincial Characteristic Key Laboratory of System Optimization and Scientific Computation,Guizhou University,Guiyang 550025,China)

机构地区：[1]贵州大学大数据与信息工程学院,贵阳550025 [2]贵州大学贵州省系统优化与科学计算特色重点实验室,贵阳550025

出　　处：《智能计算机与应用》2023年第2期179-186,共8页Intelligent Computer and Applications

基　　金：国家自然科学基金(62063002)。

摘　　要：针对YOLOv3存在行人遮挡、漏检下检测精度低、漏检率高的问题,本文提出一种基于通道注意力及空间金字塔的改进型YOLOv3。将通道注意力机制嵌入残差网络中,强化关键信息的特征提取;利用能增大感受野的空间金字塔融合多尺度特征图,增强相互遮挡目标的特征提取能力;利用改进型非极大抑制模块消除冗余预测框,避免重叠目标漏检。比较性的数值实验表明,相较于YOLOv3,改进型YOLOv3的检测准确率及综合评估指标值分别提高了9.33%和5.01%,且行人检测的鲁棒性和泛化能力更强。Aiming at the problems of pedestrian occlusion,low detection accuracy and high misdetection rate in YOLOv3,an improved YOLOv3 is proposed based on channel attention and spatial pyramid.In the design of model,the channel attention mechanism is embedded into the residual network to strengthen the feature extraction of key information.The spatial pyramid that can enlarge the receptive field is used to fuse multi-scale feature maps in order to enhance the detection of mutually occluded targets.The non-maximum suppression model is improved and used to eliminate redundant prediction frames and avoid detection failures of overlapping targets.The comparative experiments have validated that compared with YOLOv3,the detection accuracy and comprehensive evaluation index value of the improved YOLOv3 are increased by 9.33%and 5.01%respectively,and meanwhile the robustness and generalization abilities of pedestrian detection are stronger.

关 键 词：特征提取 YOLOv3 注意力机制 空间金字塔 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]