摘要
针对卷积神经网络模型体积大、运算量高,在体积小、资源有限的嵌入式平台上运行效率低,而现有轻量化模型无法兼顾检测速度和检测精度的问题,提出了一种基于Ghost模块的YOLO目标识别算法GS-YOLO。以YOLOv4模型为基础,基于Ghost模块重构目标识别网络,减少模型参数与卷积运算量,提升目标识别速率;通过融入多个空间金字塔池化模块优化目标识别精度;利用通道剪枝极限压缩方法剔除冗余参数,进一步减小模型体积与计算量;利用微调技术提升剪枝后模型的精度。实验结果表明:在自主构建的测试集和相同的测试环境下,与YOLOv4相比,GS-YOLO将YOLOv4模型体积压缩96%,浮点型计算量减少91.2%,预测速度提升2.9倍,压缩后模型识别精度达到87.63%,精度仅损失2.43%。
The convolutional neural network model with large volume and computation amount cannot be run by the embedded platform with small volume and limited resources,and the existing lightweight models are unable to juggle detection speed and accuracy.An object recognition algorithm of YOLO based on Ghost module(GS-YOLO)was proposed in this study.Following YOLOv4 model,the object recognition network was reconstructed based on Ghost module to reduce model parameters and convolution operations and improve object recognition rate.Multiple spatial pyramid pooling modules were incorporated to optimize the identification model and improve the accuracy of model identification.The channel pruning limit compression method was adopted to eliminate the redundant parameters and further reduce the model volume and calculation.The accuracy of pruning model was improved by fine tuning technique.Experimental data show that for the self-built test set and the same test environment,compared with YOLOv4 object recognition algorithm,GS-YOLO algorithm reduces the volume of YOLOv4 model by 96%,reduces the amount of floating point calculation by 91.2%,and increases the prediction speed by 2.9 times.After compression,the model recognition accuracy achieves 87.63%,and the accuracy loss is only 2.43%.
作者
符惠桐
王鹏
李晓艳
吕志刚
邸若海
FU Huitong;WANG Peng;LI Xiaoyan;Lü Zhigang;DI Ruohai(School of Electronic and Information Engineering,Xi’an Technological University,Xi’an 710021,China)
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2021年第7期124-131,共8页
Journal of Xi'an Jiaotong University
基金
国家自然科学基金资助项目(62031021)
西安市科技计划资助项目(2020KJRC0037)
西安工业大学校长基金面上培育项目(XGPY200217)。。
