A Lightweight Tomato Fruit Detection Algorithm Based on Improved Real-Time Detection Transformer
Lu Chengfang
Cui Yanrong
Hu Ronghua
Wang Haoyu
Chen Pengxiang
Abstract：Targeted at the problem of high density of targets,varying shapes and diverse distributions leading to difficult recognition in intelligent picking tomato fruits in complex environments,and also to solve deployment difficulty caused by the massive computational requirements of Transformer architectures,we pro-posed an enhanced version of the Real-Time Detection Transformer(RT-DETR),termed SPC-DETR.Firstly,multi-scale tomato fruit images in complex environments were field-taken and collected from publicly available datasets.After annotation and data augmentation,a tomato fruit dataset containing 3 398 images was construc-ted.Secondly,using RT-DETR-R18 as baseline model,improvements were made as follows:the StarNet was chosen as the benchmark network due to its low computational complexity and high detection accuracy,and an inverted residual mobile block(iRMB)was introduced to construct iR-StarNet as backbone network,impro-ving the feature extraction capability while maintaining a lightweight backbone design.An improved parallel di-lated convolution block(EMA-PDC)were designed through an efficient multi-scale attention mechanism with cross-space learning,which could reduce computational load while expanding the model's receptive field,thereby improving the accuracy in multi-scale representation.The neck network was improved by CGA-Fusion,a content-guided attention fusion scheme based on DEA-Net,to further improved the model's representational ability.The experimental results showed that SPC-DETR achieved the precision,recall and mean average pre-cision(mAP50)of 88.0%,83.7%and 90.2%respectively,which increased by 2.0,0.3 and 1.7 percentage points compared to the baseline model RT-DETR-R18,while the model weight,parameter count and floating-point operation was 25.2 MB、12.9 M、34.3 GFLOPs respectively,decreasing 34.72%,36.14%and 41.47%compared with RT-DETR-R18,outperforming most mainstream YOLO and DETR series models.These find-ings could provide references for improving the tomato fruit detection accuracy of intelligent picking robot in complex environments.
Keywords：Tomato fruit recognitionTransformerRT-DETRLightweightReal-time target detectionDeep learning
Publication Date：2026-01-30
Online Publishing Date：2026-05-22(First online date of this platform, not the publication date of the document)
Pages：14( 150-163 )