基于改进YOLOv5s的水果图像识别  

Fruit image recognition based on improved YOLOv5s

作　　者：罗荔豪 欧鸥[1] 赵伟[1] 黄元 刘学虎 LUO Li-hao;OU Ou;ZHAO Wei;HUANG Yuan;LIU Xue-hu(College of Computer and Network Security(Oxford Brookes College),Chengdu University of Technology,Chengdu 610059,China)

机构地区：[1]成都理工大学计算机与网络安全学院(牛津布鲁克斯学院),成都610059

出　　处：《信息技术》2023年第11期28-34,40,共8页Information Technology

基　　金：国家重点研发计划资助项目(2018YFF01013304)。

摘　　要：针对已有的水果识别模型在复杂场景下识别准确率低、推理耗时长、对硬件要求高等问题,提出了一种适用于复杂场景下轻量级的水果识别模型YOLOv5s_CB_CA。在YOLOv5s目标检测算法的基础上,对主干特征提取网络Backbone引入卷积注意力机制CBAM,丰富空间和通道维度的特征信息;并采用特征图上采样CARAFE替换原始上采样,减少计算量、提高识别速度。将改进算法与YOLOv5s算法在水果图像数据集作对比。实验结果表明:YOLOv5s_CB_CA算法平均精度(mAP)达到了96.5%、召回率(Recall)达到了93.5%,使模型体积缩小了约14%。YOLOv5s_CB_CA算法提高了检测精度和召回率、缩小了模型体积,证明了该算法的有效性。To solve the problems of low recognition accuracy,long reasoning time and high hardware requirements of existing fruit recognition models in complex scenes,a lightweight fruit recognition model YOLOv5s_CB_CA was proposed.On the basis of YOLOv5s object detection algorithm,Convolution Attention Mechanism CBAM is introduced into Backbone feature extraction network to enrich feature information of spatial and channel dimensions.The original up-sampling is replaced by up-sampling CARAFE,which reduces computation and improves recognition speed.The improved algorithm is compared with YOLOv5s algorithm in fruit image data set.Experiment results show that the YOLOv5s_CB_CA algorithm achieves 96.5%average accuracy(mAP)and 93.5%Recall,and reduces the model volume by 14%.The YOLOv5s_CB_CA algorithm improves the detection accuracy and recall rate,and reduces the model size,which proves the effectiveness of the algorithm.

关 键 词：水果识别 目标检测 YOLOv5s算法 卷积注意力机制CBAM 轻量级算子CARAFE 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]