Jiang Haoyu, Hu Fei, Fu Xiuqing, Chen Cairong, Wang Chen, Tian Luxu, Shi Yuran
College of Artificial Intelligence, Nanjing Agricultural University, Nanjing, China.
College of Engineering, Nanjing Agricultural University, Nanjing, China.
Front Plant Sci. 2023 Sep 28;14:1257947. doi: 10.3389/fpls.2023.1257947. eCollection 2023.
Drought stress has become an important factor affecting global food production. Screening and breeding new varieties of peas (Pisum sativum L.) for drought-tolerant is of critical importance to ensure sustainable agricultural production and global food security. Germination rate and germination index are important indicators of seed germination vigor, and the level of germination vigor of pea seeds directly affects their yield and quality. The traditional manual germination detection can hardly meet the demand of full-time sequence nondestructive detection. We propose YOLOv8-Peas, an improved YOLOv8-n based method for the detection of pea germination vigor.
We constructed a pea germination dataset and used multiple data augmentation methods to improve the robustness of the model in real-world scenarios. By introducing the C2f-Ghost structure and depth-separable convolution, the model computational complexity is reduced and the model size is compressed. In addition, the original detector head is replaced by the self-designed PDetect detector head, which significantly improves the computational efficiency of the model. The Coordinate Attention (CA) mechanism is added to the backbone network to enhance the model's ability to localize and extract features from critical regions. The neck used a lightweight Content-Aware ReAssembly of FEatures (CARAFE) upsampling operator to capture and retain detailed features at low levels. The Adam optimizer is used to improve the model's learning ability in complex parameter spaces, thus improving the model's detection performance.
The experimental results showed that the Params, FLOPs, and Weight Size of YOLOv8-Peas were 1.17M, 3.2G, and 2.7MB, respectively, which decreased by 61.2%, 61%, and 56.5% compared with the original YOLOv8-n. The mAP of YOLOv8-Peas was on par with that of YOLOv8-n, reaching 98.7%, and achieved a detection speed of 116.2FPS. We used PEG6000 to simulate different drought environments and YOLOv8-Peas to analyze and quantify the germination vigor of different genotypes of peas, and screened for the best drought-resistant pea varieties.
Our model effectively reduces deployment costs, improves detection efficiency, and provides a scientific theoretical basis for drought-resistant genotype screening in pea.
干旱胁迫已成为影响全球粮食生产的重要因素。筛选和培育耐旱豌豆（Pisum sativum L.）新品种对于确保农业可持续生产和全球粮食安全至关重要。发芽率和发芽指数是种子发芽活力的重要指标，豌豆种子的发芽活力水平直接影响其产量和品质。传统的人工发芽检测很难满足全时序无损检测的需求。我们提出了YOLOv8-Peas，一种基于YOLOv8-n改进的豌豆发芽活力检测方法。
我们构建了一个豌豆发芽数据集，并使用多种数据增强方法来提高模型在实际场景中的鲁棒性。通过引入C2f-Ghost结构和深度可分离卷积，降低了模型的计算复杂度并压缩了模型大小。此外，用自行设计的PDetect检测头替换原始检测头，显著提高了模型的计算效率。在骨干网络中添加坐标注意力（CA）机制，以增强模型从关键区域定位和提取特征的能力。颈部使用轻量级的特征内容感知重组（CARAFE）上采样算子来捕获和保留低层次的详细特征。使用Adam优化器来提高模型在复杂参数空间中的学习能力，从而提高模型的检测性能。
实验结果表明，YOLOv8-Peas的参数（Params）、浮点运算次数（FLOPs）和权重大小（Weight Size）分别为117万、32亿和2.7MB，与原始的YOLOv8-n相比分别下降了61.2%、61%和56.5%。YOLOv8-Peas的平均精度均值（mAP）与YOLOv8-n相当，达到98.7%，并实现了1******
