YOLOC-tiny: a generalized lightweight real-time detection model for multiripeness fruits of large non-green-ripe citrus in unstructured environments

• DOI: 10.3389/fpls.2024.1415006
• Publication Date: 2024-07-05T05:14:16Z
• Authors: Zuoliang Tang; Lijia Xu; Haoyang Li; Mingyou Chen; Xiaoshi Shi; Long Zhou; Yuchao Wang; Zhijun Wu; Yongpeng Zhao; Kun Ruan; Yong He; Wei Ma; Ning Yang; Lufeng Luo; Yunqiao Qiu
• Keywords: Orchard

This study addresses the challenges of low detection precision and limited generalization across various ripeness levels varieties for large non-green-ripe citrus fruits in complex scenarios. We present a high-precision lightweight model, YOLOC-tiny, built upon YOLOv7, which utilizes EfficientNet-B0 as feature extraction backbone network. To augment sensing capabilities improve accuracy, we embed spatial channel composite attention mechanism, convolutional block module (CBAM), into head’s efficient aggregation Additionally, introduce an adaptive complete intersection over union regression loss function, designed by integrating phenotypic features citrus, to mitigate impact data noise efficiently calculate loss. Finally, layer-based magnitude pruning strategy is employed further eliminate redundant connections parameters model. Targeting three types widely planted Sichuan Province—navel orange, Ehime Jelly Harumi tangerine—YOLOC-tiny achieves impressive mean average (mAP) 83.0%, surpassing most other state-of-the-art (SOTA) detectors same class. Compared with YOLOv7 YOLOv8x, its mAP improved 1.7% 1.9%, respectively, parameter count only 4.2M. In picking robot deployment applications, YOLOC-tiny attains accuracy 92.8% at rate 59 frames per second. provides theoretical foundation technical reference upgrading optimizing low-computing-power ground-based robots, such those used fruit orchard inspection.