Authors

Abstract

To address the issues of low detection accuracy, slow speed, and large parameter size in detecting fresh table grapes in natural orchard environments, this study proposes an improved grape detection model based on YOLOv8n, termed YOLOGPnet. The model replaces the C2f module with a Squeeze-and-Excitation Network V2 (SENetV2) to enhance gradient flow through more branched cross-layer connections, thereby improving detection accuracy. Additionally, the Spatial Pyramid Pooling with Enhanced Local Attention Network (SPPELAN) substitutes the SPPF module, enhancing its ability to capture multi-scale information of the target fruits. The introduction of the Focaler-IoU loss function, along with different weight adjustment mechanisms, further improves the precision of bounding box regression in object detection. After comparing with multiple algorithms, the experimental results show that YOLOGPnet achieves an accuracy of 93.6% and [email protected] of 96.8%, which represents an improvement of 3.5 and 1.6 percentage points over the baseline model YOLOv8n, respectively. The model's computational load, parameter count, and weight file size are 6.8 Gflops, 2.1 M, and 4.36 MB, respectively. The detection time per image is 12.5 ms, showing reductions of 21.84%, 33.13%, 30.79%, and 25.60% compared to YOLOv8n. Additionally, comparisons with YOLOv5n and YOLOv7-tiny in the same parameters reveal accuracy improvements of 0.7% and 1.9%, respectively, with other parameters also showing varying degrees of enhancement. This study offers a solution for accurate and rapid detection of table grapes in natural orchard environments for intelligent grape harvesting equipment.

Abstract in Chinese

针对自然果园环境下鲜食葡萄的检测精度低、速度慢、参数量较大等问题，本研究提出了一种基于改进YOLOv8n的葡萄检测模型（YOLOGPnet）。该模型使用压缩与激励网络（Squeeze-and-Excitation Network V2，SENetV2）替换了C2f模块，通过更多的分支跨层连接使梯度流更加丰富，提高模型的检测精度；并将SPPF模块替换为增强局部注意力的空间金字塔池化网络（Spatial Pyramid Pooling with Enhanced Local Attention Network，SPPELAN），提升了网络捕捉目标果实的多尺度信息的能力；通过使用Focaler-IoU损失函数，和引入不同的权重调整机制提高了目标检测中的边界框回归精度问题。与多种算法进行了比较后，试验结果表明，YOLOGPnet的精确度和[email protected]分别为93.6%、96.8%，相较于基准模型YOLOv8n的原始水平分别提高了3.5和1.6个百分点。该模型的计算量、参数量和权重文件大小分别为6.8 Gflops、2.1 M和4.36 MB，单幅图像检测耗时为12.5 ms，相较于YOLOv8n，分别降低了21.84%、33.13%、30.79%和25.60%，同时将改进后的模型与YOLOv5n和YOLOV7-tiny也进行了上述参数的对比，在精度方面分别提升了0.7％和1.9％，在其他参数方面也都存在不同程度的提高。该研究为智能化葡萄采摘装备在自然果园环境下准确且快速地检测鲜食葡萄提供了一种解决方案。