基于改进YOLO v8n的田间玉米杂草快速识别方法

Corn and Weeds Faster Identification Method Based on Improved YOLO v8n in Field

摘要:

玉米是东北黑土区主要粮食作物，苗期常受伴生杂草威胁，化学和机械除草作为重要的中耕管理手段，其实施效果对黑土地保护意义重大，杂草防治需以高效精准的杂草识别技术为基础。为实现玉米田间多行作物环境下杂草高效精准识别，以东北黑土区2～5叶期玉米幼苗及其早期伴生杂草为对象，提出了一种基于改进YOLO v8n的玉米杂草高效精准识别算法。通过将YOLO v8n模型主干网络替换为GhostNet轻量特征提取网络，简化了模型特征提取流程并缩减了模型参数量;通过引入优化的加权双向特征金字塔结构（Bidirectional feature pyramid network，BiFPN），并在颈部网络最后一层引入轻量注意力机制（Triplet Attention），增强了模型信息获取能力，提高了模型泛化能力;通过将损失函数替换为Wise?IoU v3，进一步提高了模型精度。试验结果表明，改进YOLO v8n平均精度均值达97.8%，相较于原始模型，改进模型参数量降低37.90%，检测速度、平均精度均值、精确率和召回率分别提高32.10%、0.65个百分点、1.85个百分点和1.40个百分点。与主流检测模型相比，检测速度相较于YOLO v5n、YOLO v6n、Faster R?CNN、YOLO v9t和YOLO v10n分别提高26.38%、33.80%、591.89%、96.58%和53.93%。该方法满足除草机器人等移动端部署要求，为东北黑土地保护性耕作中耕管理环节智能化机械除草和精准对靶施药提供了技术支持。

Abstract:

Corn is one of the three major grain crops in China, and precise application of weeds at the seedling stage is crucial to ensure its yield and quality. Weeds in the seedling stage compete with corn for sunlight, water, and nutrients, seriously affecting the normal growth of corn and even leading to yield reduction or death. Therefore, accurate and fast weed identification is the basis for precise weed control. Existing weed recognition in field environments is mostly based on small field?of?view image datasets, and few studies related to corn weed recognition in field multi?row environments are seen. In the field multi?row environment, because a single image data can cover multiple rows of crops at the same time, the number of target objects is dense, and the recognition range of weeds is larger in a single detection. However, because a variety of weeds are common in the seedling stage of corn, and the weeds and crops overlap each other, the image data in the multi?row environment is more complex, and it is more difficult to control the detection accuracy and computational volume of the conventional model. To address the above problems, northeast seedling corn was taken as the research object. The corn seedling and weed datasets used for model training and containing both single and multi?targets were collected, and data enhancement operations were performed on the datasets. The specific main content was to address the problem of insufficient target detection speed in the process of precision application and weed control, which led to the deviation of target application. A field corn weed identification method with an improved YOLO v8n model was proposed by using a northeastern corn field as the research object. By collecting pictures of corn seedlings and weeds in the field under the range of multiple rows as a training set, which included single weeds, multiple weeds of multiple species, and the coexistence of corn seedlings and weeds, the backbone network of the YOLO v8 model was replaced by the GhostNet lightweight feature extraction network to simplify the feature extraction process and reduce the number of model parameters. The weighted bidirectional feature pyramid network (BiFPN) was introduced to enhance the information acquisition ability of the model. Triplet Attention, a lightweight attention mechanism, was introduced into the neck network of the model, and the loss function was replaced with Wise?IoU v3 to improve the generalization ability and fine reading of the model. The experimental results showed that compared with the original model, the improved model parameter count was reduced by 37.90%, the detection speed was increased by 32.10%, the mean average precision (mAP) was improved by 0.65 percentage points, the precision was improved by 1.85 percentage points, and the recall was improved by 1.40 percentage points. The research results can provide a basis for decision?making in herbicide reduction application.