Detection for Harvesting with an Active Illumination Camera System and DUTU2-Net+

Qinghui Pan; Jie Lian; Yadong Zhao; Chaochao Qiu; Dong Wang

IROS 20250 citations

Detection for Harvesting with an Active Illumination Camera System and DUTU2-Net+

Qinghui Pan, Jie Lian, Yadong Zhao, Chaochao Qiu, Dong Wang

Abstract

Robots operating in agricultural environments require a robust, fast perception system to accurately identify picking points. This paper proposed a lightweight method for detecting sweet pepper peduncles, which uses an active illumination camera system and DUTU2-Net+ to achieve efficient and accurate peduncle detection. The camera system is used to overcome the influence of ambient light through flash and no-flash (FNF) image pairs, achieving more robust color detection and quickly locating the peduncle’s region of interest (ROI). The improved DUTU2-Net+ is used accurately for peduncle ROI detection. It uses an encoder with depthwise separable convolution (DSC), dilated convolution, and a feature enhancement structure with a triple attention module (TAM) to reduce the computational load and parameters while ensuring detection accuracy. Experimental results show that the proposed method can effectively identify the position of the peduncle. The DUTU2-Net+ model achieves an average absolute error of 0.002, a maximum F<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> score of 0.992, a frame rate of 36.3 FPS, and a model size of 6.9 MB. The source code is available at https://gitee.com/rosdx/detection-for-harvesting.git.

BibTeX

@inproceedings{iros2025_detectionforharv,
  title = {Detection for Harvesting with an Active Illumination Camera System and DUTU2-Net+},
  author = {Qinghui Pan and Jie Lian and Yadong Zhao and Chaochao Qiu and Dong Wang},
  booktitle = {IROS 2025},
  year = {2025}
}