基于LC一YOLO的自然场景荔枝果实实时检测算法

打开文本图片集

中图分类号：S667.1；TP391.4 文献标识码：A 文章编号：2095-5553（2026）03-0059-09

DOI：10.13733/j.jcam.issn.2095-5553.2026.03.009

Abstract：Manuallycheepickingisineficientandlabor-intensive，whilemachine-vision-basedautomateddetection andharvestingoferbothefficiencyandsafety.However，real-worldchallengessuchastheneedfor high real-time performance，diffcultyin detecting small targets，complex background interference，and fruit occlusion，make automatic detection particularly challenging.To address these issues，this paper proposes an LC—YOLO（Litchi YOLO）detection algorithm for lychee fruit recognition.Building on the YOLOvll architecture，an AKConv module was designedand integratedto enhance featureextractioncapabilities whilereducingmodel complexity.AGuided Attention；Module（GAM）was constructed to enhance multi-scale feature fusion capabilitiesandsuppress background noise.In addition，the MPDIoU loss function wasadopted toimprove localization accuracy indense fruit scenarios.Experimental resultsshow that LC—YOLO significantlyoutperformsmainstreamobjectdetection modelsin detectionaccuracy，model lightweight design，andreal-timeperformance.It achieves arecall rateof （20 95.6% and an mAP of 94.5% ，with real-time performance reaching 96.8 frames per second. Additionally，on the resource-constrained Jetson Nano embedded platform，LC—YOLOachieves an inference speedof 26.4 frames per second withonly 7.9W of power consumption，demonstrating excellent cross-platform adaptability and robustness.

Keywords：litchi； fruit detection；natural scenes；lightweight model；attention mechanism；loss function

0 引言

中国是全球最大的荔枝生产国，2023年总产量达3924.3kt ，并呈现持续增长趋势[1]。（剩余13547字）