江西农业大学学报2023，Vol.45Issue(6)：1528-1542,15.DOI:10.13836/j.jjau.2023140
基于改进YOLOv7-Tiny的成熟草莓识别模型研究
Research on Ripe-strawberry Recognition Model Based on Improved YOLOv7-Tiny
摘要
Abstract
[Objective]Rapid and accurate recognition of ripe strawberries is a crucial technology for efficient mechanical harvesting.Aiming at the problems of fruit stacking,leaf occlusion and small targets in strawberry growth environment,a mature strawberry recognition model based on improved YOLOv7-Tiny is proposed.[Method]The model builds on the YOLOv7-Tiny model.The LeakeyReLU activation function of the CBL convolutional block in the YOLOv7-Tiny model backbone network was replaced with the SiLU function to improve the nonlinear fitting degree and feature learning ability of the model.In order to reduce the parameter quantity and calculation amount of the model,realize the lightweight of the model and improve the recognition speed,a lightweight RepGhost network is introduced.The C3 module was added to the small object layer of the YOLOv7-Tiny model.The number of model parameters was reduced and the network depth of the model was increased.The information extraction ability of the model for small-size objects was enhanced and the accuracy of the model in recognizing overlapped strawberries and small object strawberries was improved.Recognition speed was further enhanced.With the strawberry as the test sample,the improved YOLOv7-Tiny model was tested to verify the performance.[Result]Compared with the YOLOv7-Tiny model,the improved YOLOv7-Tiny model had fast convergence speed,small and stable fluctuation range of the loss curve after model fitting,small loss value of training,and good robustness of the model.The results of the comparative tests before and after the improvement showed that the number of parameters of the improved YOLOv7-Tiny model was reduced by 26.9%.The model computation was reduced by 55.4%.The model recognition speed was increased by 26.3%.The mean accuracy of recognition(mAP)was 89.8%.The improved YOLOv7-Tiny model improved the performance in all aspects when compared with the YOLOv7-Tiny model.The ablation test verified that the SiLU activation function,RepGhostNet,and C3 module effectively improved the recognition speed and accuracy of the YOLOv7-Tiny model.The improved YOLOv7-Tiny model was tested for performance comparison with SSD,Faster RCNN,YOLOv3,YOLOv4,and YOLOv5s models.The results showed that the F1 score of the improved YOLOv7-Tiny model was 0.87.The improved YOLOv7-Tiny model had a higher F1 score than other deep learning models.The mAP of the YOLOv7-Tiny model increased by 14.2%,1.52%,3.15%,3.01%,and 2.6%,respectively.The recognition speed increased by 79.3%,92.9%,80.4%,58.8%,and 69.6%,respectively.The number of parameters decreased by 90%,89.7%,95%,47.8%and 14.6%,respectively.[Conclusion]The improved YOLOv7-Tiny model is characterized by fast recognition,high recognition accuracy,and lightweight.The accuracy of the improved YOLOv7-Tiny model in recognizing overlapped ripe strawberries is significantly improved.The study provides technical supports for the efficient recognition of ripe strawberries.关键词
深度学习/草莓识别/YOLOv7-Tiny/轻量化/小目标/RepGhostKey words
deep learning/strawberry recognition/YOLOv7-Tiny/lightweighting/small target/RepGhost分类
农业科技引用本文复制引用
汤泽政,伍奕桦,徐新明,郭建政,童成彪..基于改进YOLOv7-Tiny的成熟草莓识别模型研究[J].江西农业大学学报,2023,45(6):1528-1542,15.基金项目
湖南省重点研发计划项目(2022NK2028) Project supported by the Hunan Provincial Key Research and Development Project(2022NK2028) 湖南省教育厅科学研究项目(21C01260)同时对本研究给予了资助,谨致谢意! （2022NK2028）
