基于改进YOLOv7的水果目标检测方法  

Fruit Target Detection Method based on Improved YOLOv7

作　　者：刘麒[1] 李奎东 常广良 王影[1] LIU Qi;LI Kuidong;CHANG Guangliang;WANG Ying(School of Information and Control Engineering,Jilin Institute of Chemical Technology,Jilin City 132022,China;Kesshida(Changchun)Automobile Electrical Appliance Co.,LTD.,Changchun 130031,China)

机构地区：[1]吉林化工学院信息与控制工程学院,吉林吉林132022 [2]科世达(长春)汽车电器有限公司,吉林长春130031

出　　处：《吉林化工学院学报》2024年第7期12-17,共6页Journal of Jilin Institute of Chemical Technology

基　　金：吉林市科技成果项目(20190502118);吉林化工学院科研项目(2018064);吉林化工学院重大科技项目(2018017)

摘　　要：目前水果种植广泛,但水果分类工作大多数是人工完成,耗费了大量人工成本,少数机器识别也有速度慢、准确率低等问题。针对此类目标检测识别效率较低等问题,提出了一种基于改进YOLOv7算法的水果目标检测算法,通过引入ECA注意力机制,增强通道维度的关联性,提升了模型的表达能力和学习效果;使用PConv代替了部分的卷积结构,同时削减冗余计算和内存访问,更有效地提取空间特征;损失函数使用了MPDIoU,针对传统IoU增强了梯度可导性,便于进行图像分割任务的训练和优化。改进的YOLOv7算法在精度和平均精度均值方面分别提升了4%和3%,能准确识别出水果。At present,fruits are widely planted,but most of the fruit classification is done manually,which consumes a lot of labor costs,and a few machine recognition also has problems such as slow speed and low accuracy.To solve the problem of low efficiency of target detection and recognition,a fruit target detection algorithm based on improved YOLOv7 algorithm was proposed.By introducing ECA attention mechanism,the relevance of channel dimensions was enhanced,and the expression ability and learning effect of the model were improved;PConv was used instead of partial convolution structure,and redundant computing and memory access were reduced to extract spatial features more effectively;The loss function uses MPDIoU,which enhances the gradient differentiability for traditional IoU,facilitating the training and optimization of image segmentation tasks.The improved YOLOv7 algorithm can accurately recognize fruits by improving the precision and average accuracy by 4%and 3%respectively.

关 键 词：水果识别 目标检测 YOLOv7 MPDIoU 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]