融合CBAM和DCNv3的YOLOv7模型果蔬检测分类研究

打开文本图片集

中图分类号：S24；TP391.41 文献标识码：A 文章编号：2095-5553（2026）02-0292-08

Abstract： In order to improve the sorting speed of fruits and vegetables and reduce labor costs，an improved YOLOv7 fruitandvegetable detectionandclasification model combining CBAMandDCNv3was proposed.The proposed approach incorporates three key innovations.First，the CBAMattention mechanism wasused in the backbone network to suppress irrlevant featureswhileenhancing keyfeatureextraction.Second，adeformableconvolutional neuralnetwork wasaddedbefore the SpatialPyramidPoling（SPP）layer toenhancethemulti-scalefeature fusionof the model. Finally，it was testedonacustom dataset of tencommon fruitsandvegetablesand the public VOC dataset for fruit classification and target detection. On the custom dataset，the recall rate of the model reached 95.0% ，the recognition accuracy reached 95.6% ，and the mAP@0.5 reached 97.2% . On the VOC dataset，the recall rate of the model reached 99.30% ，the recognition accuracy reached 98.10% ，andthe mAP@0.5 reached 98.50% .The experimental results showed that theproposed model achievedsuperior performance in terms of recognition acuracy，robustness，and generalization capabilities compared with other mainstream vegetable and fruit recognition algorithms.

Keywords：fruit and vegetable detection；multi-scale feature；deformable convolution；attention module

0 引言

果蔬富含丰富的维生素，需求量逐年增加，传统的人工识别和分类已经很难满足果蔬行业日益增长需求。（剩余10628字）