基于SAW-YOLO v8n的葡萄幼果轻量化检测方法

Lightweight Detection Method for Young Grape Cluster Fruits Based on SAW-YOLO v8n

摘要:

葡萄簇幼果果实受背景色、遮挡和光照变化的影响，检测难度大。为了实现对背景色、遮挡和光照变化具有鲁棒性的葡萄簇幼果检测，提出了一种融合随机注意力机制（Shuffle attention，SA）的改进YOLO v8n模型（SAW-YOLO v8n)。通过在YOLO v8n模型的Neck结构中融入SA机制，增强网络多尺度特征融合能力，提升检测目标的特征信息表示，并抑制其他无关信息，提高检测网络检测精度，在不明显增加网络深度和内存开销的情况下，实现了葡萄簇幼果的高效准确检测;采用基于动态非单调聚焦机制的损失（Wise intersection over union loss，Wise-IoU Loss）作为边界框回归损失函数，加速网络收敛并进一步提高模型的准确率。构建了葡萄簇幼果的数据集GGrape，该数据集由3 780幅复杂场景下的葡萄簇幼果图像及对应标注文件组成。通过该数据集对SAW-YOLO v8n模型进行训练和测试。测试结果表明，基于SAW-YOLO v8n的葡萄簇幼果检测算法的精度（Precision，P）、召回率（Recall，R）、平均精度均值（Mean average precision，mAP）和F1值分别为92.80%、91.30%、96.10%和92.04%，检测速度为140.85 f/s，模型内存占用量为6.20 MB。与SSD、YOLO v5s、YOLO v6n、YOLO v7-tiny、YOLO v8n等5个轻量化模型相比，其mAP值分别提高16.06%、1.05%、1.48%、0.84%、0.73%，F1值分别提高24.85%、1.43%、1.43%、1.09%、1.60%，模型内存占用量分别降低93.16%、56.94%、37.63%、47.00%、0，是所有模型中最小的，具有明显的轻量化、高精度优势。讨论了不同遮挡程度和光照条件的葡萄幼果检测，结果表明，基于SAW-YOLO v8n的葡萄幼果检测方法能适应不同遮挡和光照变化，具有良好的鲁棒性。结果表明，SAW-YOLO v8n不仅能满足对葡萄簇幼果检测的高精度、高速度、轻量化的要求，且具有较强的鲁棒性和实时性。

Abstract:

The detection of young grape cluster fruits is challenging due to the influence of background color, occlusion, and lighting variations. To achieve robust detection of young grape cluster fruits for the varying conditions, an improved YOLO v8n model that integrated shuffle attention (SA) mechanism was proposed in the work. By incorporating SA mechanism into the Neck network of the YOLO v8n model, the multi-scale feature fusion ability of the network was enhanced, the feature information representation of the detection target was improved, and other irrelevant information was suppressed, improving the accuracy of the detection network, which achieved efficient and accurate detection of young grape cluster fruits without significantly increasing network depth and memory overhead. Wise intersection over union loss (Wise-IoU Loss) with the dynamic nonmonotonic focusing mechanism was taken as the bounding box regression loss function, to accelerate the network convergence for the better detection accuracy of the model. Herein, a Grape dataset was constructed, which comprised 3 780 images of young grape cluster fruits in complex scenarios along with corresponding annotation files. Training and testing results of the SAW-YOLO v8n model on this dataset showed that the precision (P), recall (R), mean average precision (mAP), and F1 score of the young grape cluster fruit detection algorithm based on SAW-YOLO v8n were 92.80%, 91.30%, 96.10%, and 92.04%, respectively, where the detection speed was 140.85 f/s, and the model size was only 6.20 MB. Compared with that of SSD, YOLO v5s, YOLO v6n, YOLO v7-tiny, and YOLO v8n, the mAP was increased by 16.06%, 1.05%, 1.48%, 0.84% and 0.73%, respectively, and F1 scores were increased by 24.85%, 1.43%, 1.43%, 1.09% and 1.60%, respectively, and the model weights were reduced by 93.16%, 56.94%, 37.63%, 47.00%, and 0, respectively, which was the smallest among all models and had obvious advantages in lightweight and high accuracy. Moreover, the young grape cluster fruits detection with different degrees of occlusion and lighting conditions were also explored, and the result showed that the young grape cluster fruit detection method based on SAW-YOLO v8n can adapt to different occlusion and lighting changes, and had good robustness. In summary, SAW-YOLO v8n not only met the requirements of high-precision, high-speed, and lightweight detection of young grape cluster fruits, but also had strong robustness and real-time performance.