现代电子技术2025，Vol.48Issue(10)：85-91,7.DOI:10.16652/j.issn.1004-373x.2025.10.014
YOLOv7-VSS轻量化橘瓣外观检测模型
YOLOv7-VSS lightweight orange segment appearance detection model
喻擎苍 1邱锐 1傅林杰 1谢淼 1孙树森1
作者信息
- 1. 浙江理工大学 计算机科学与技术学院,浙江 杭州 310018
- 折叠
摘要
Abstract
In allusion to the problems of slow speed and low accuracy of orange petal appearance detection in canned citrus production,as well as the higher parameter count of mainstream detection models,a lightweight orange segment appearance detection model,YOLOv7-VSS,is proposed.In the model,an improved EfficientViT network is introduced by using the HardSwish activation function as the backbone.The mapping similarity between different detection heads is reduced by inputting features at different levels,which alleviates redundant calculations,and enhances the network′s feature extraction capability by means of cascaded group attention mechanism.A slim-neck module that fuses the properties of standard convolution and deep separable convolution is referenced to reduce the size of the model while maintaining high accuracy.In order to further reduce the model size and speed up inference speed,SPPCSPC is replaced with the SPPF structure.In order to align with the positional characteristics of orange segments in the dataset,the MPDIoU loss function is used to improve the regression accuracy of the predicted bounding boxes.The experimental results show that the proposed orange segment appearance detection model is 63.81%smaller in size compared to YOLOv7,while realizing a detection accuracy of 96.57%.After deployment and testing on the Jetson Orin Nano,the balance between model size and detection accuracy is significantly improved compared to similar methods,meeting the requirements of the canned citrus production line.关键词
橘瓣外观检测/YOLOv7/轻量化/EfficientViT/GSConv/Hard-Swish/MPDIoUKey words
orange segment appearance detection/YOLOv7/lightweight/EfficientViT/GSConv/Hard-Swish/MPDIoU分类
信息技术与安全科学引用本文复制引用
喻擎苍,邱锐,傅林杰,谢淼,孙树森..YOLOv7-VSS轻量化橘瓣外观检测模型[J].现代电子技术,2025,48(10):85-91,7.
