基于Real-ESRGAN预处理的YOLOv7菜品识别方法

打开文本图片集

中图分类号：TP391.4；0235 文献标识码：A 文章编号：2096-4706（2026）06-0055-09

YOLOv7 Dish Recognition Method Based on Real-ESRGAN Preprocessing

HOU Tingting'， ZHANG Hui²， BAI Yuxin²， LIU Ziwei²， LIU Wanting² （1.Basic TeachingDepartment，Anhui Business College，Wuhu 241o02，China; 2.SchoolofComputingScience，WuhuUniversity，Wuhu 241oo8，China）

Abstract： Aiming at the problems of object detection and clasification incanteen dish recognition，this paper proposes afusionalgorithm based oncomputer vision.The model is trained on aself-built canten dish dataset under the YOLOv7 environment，and various data augmentation and combinationaugmentation methods areused to evaluate model performance. SubsequentlyReal-ERGANis introduced toperform denoisingandsuper-resolutionreconstructiononimagedatatomprove theclarityandfeaturequalityofinput images.Experimentalresultsontheself-builtdatasetcontaining10ooimagesshow that compared withtheoriginalYOLOv7method withoutReal-ESRGANpreprocesing，the improvedmodelintegratingRealESRGAN increases by approximately 5% in the mAP（∅0.5 index and by approximately 3% in the stricter m ιAP@0.5：0.95 （20号 index.Among them， the performance is optimal under the high-definition ×2 condition， where mAP（∅0.5 reaches 0.98 and mAP@0.5：0.95 is approximately 0.70.The results indicate that this method effectively improves the accuracy of dish recognition andhasgood generalization performance and application potential.

Keywords： YOLOv7; Real-ESRGAN; image recognition; algorithm optimization

0 引言

由于菜品品类繁多，早期识别方法通常先提取图像特征点，再借助机器学习训练分类器，以实现高效自动识别。（剩余14009字）