【作者】 刘伟；

【导师】 洪青青；

【作者基本信息】 扬州大学 ， 软件工程， 2025， 硕士

【摘要】 在小麦育种过程中,穗数是选择优良品种的关键表型性状指标之一,直接影响品种的选育效果。准确统计穗数可以帮助育种学家筛选出高产、优质的品种,对农业生产与研究具有重要价值。卷积神经网络凭借其强大的局部特征提取能力,被广泛的应用到农业领域。由于田间复杂环境导致的背景干扰等因素,传统的卷积神经网络受限于卷积核尺寸,难以有效的捕获麦穗图像的全局上下文信息,这一定程度上制约了麦穗计数的精度。此外,目前大多数的麦穗计数方法主要使用全监督学习方式训练模型,需要大量的标记数据,耗费较高的人力成本且模型体量大,难以部署到边缘设备。针对上述问题,本研究提出了一种基于多尺度特征融合的麦穗计数网络轻量化方法,主要研究内容如下:1)卷积神经网络具有强大的局部特征提取能力,但由于卷积核感受野的限制,无法捕捉全局上下文信息,从而影响了计数性能。针对这一问题,本研究通过融合局部细节特征与全局上下文信息,设计了一种新型计数网络模型,实现从RGB图像中计数小麦麦穗。一方面,通过采用跨阶段残差结构构建卷积神经网络,以提取多尺度局部特征;另一方面,利用自注意力机制建立特征间的长距离依赖关系加强对全局上下文信息的提取能力,增强特征表征能力,从而提升模型的计数性能。本研究采用GWHD和WEDD两个麦穗检测数据集作为测试数据进行实验验证,实验结果显示模型平均绝对误差分别为3.40和5.21,显著优于现有的相关研究成果。2)基于深度学习的麦穗计数方法有了很大的改进。但当前的计数方法仍然面临着数据采集成本高昂和标注流程复杂等实际挑战。因此,针对数据样本标记繁琐,耗费大量时间的问题。本研究中,提出了一种半监督计数网络用于实现麦穗计数。本网络在均值教师的模型的基础上,利用注意力权重,有效的融合局部特征和全局上下文信息。最后通过使用多尺度膨胀卷积构建回归头和注意力生成器,以细粒度的方式处理麦穗特征图生成预测密度图和注意力图。另外,本研究引入掩码图像建模策略来促进对特征的理解能力。实验数据表明,在标注数据比例从10%增至70%的对比实验中,本文提出的半监督模型性能优于其他半监督模型。在只有30%标记数据情况下,模型取得平均绝对误差和均方根误差分别为6.71和9.02的优异表现,证明了其在有限标注数据场景下的优越性。3)深度学习网络性能的增强通常是以增加网络规模和计算复杂性为代价的。对于田间的小麦麦穗数量统计,为边缘设备设计轻量级的麦穗计数模型也具有挑战性。因此,为了解决这个问题,提出了一种轻量级的麦穗计数网络用于有效的学习全局特征和局部特征。具体来说,设计了一个多尺度的状态空间组来处理多个不同尺度的图像序列块,旨在提取不同尺度的全局特征。其次使用一个轻量化的卷积网络分支获取局部特征。最后通过全局-局部自适应融合模块,通过权重动态计算机制融合多尺度的全局特征和局部特征。实验表明,该轻量化模型仅需7.47M参数,获得的了平均绝对误差为5.50,均方根误差为7.11,FPS达到43的效果。更多还原

【Abstract】 In the process of wheat breeding,spike number is one of the key phenotypic trait indexes for selecting good varieties,which directly affects the selection effect of varieties.Accurate statistics of spike number can help breeders to select high-yielding and high-quality varieties,which is of great value to agricultural production and research.Convolutional neural networks are widely used in the field of agriculture by virtue of their powerful local feature extraction capability.Due to factors such as background interference caused by the complex environment in the field,traditional convolutional neural networks are limited by the size of the convolutional kernel,and it is difficult to effectively capture the global contextual information of the wheat ear image,which restricts the accuracy of wheat ear counting to a certain extent.In addition,most of the current wheat ear counting methods mainly use fully supervised learning to train the model,which requires a large amount of labeled data,consumes high labor costs and has a large model size,making it difficult to deploy to edge devices.Aiming at the above problems,the main research content of this study is as follows:1)While convolutional neural networks have strong local feature extraction capabilities,their sense-field limitations cause the network to capture global contextual information,which affects counting performance.To address this problem,this paper designs a novel counting network model to realize counting wheat sheaves from RGB images by fusing local detail features with global contextual information.On the one hand,a convolutional neural network is constructed by employing a cross-stage residual structure to extract multi-scale local features.On the other hand,the self-attention mechanism is used to establish long-distance dependencies between features to strengthen the ability to extract global contextual information and enhance feature characterization,thus improving the counting performance of the model.In this paper,two wheat detection datasets,GWHD and WEDD,were used as testing data for experimental validation,and the experimental results showed that the average absolute errors of the models were 3.40 and 5.21,respectively,which significantly outperforms the existing research results.2)Deep learning-based methods for counting wheat ears detection have improved greatly.But still limited by cumbersome and expensive data collection and labeling.Therefore,the problem of tedious and time-consuming labeling of data samples is addressed.In this paper,a wheat ears counting model with semi-supervised counting network is proposed.Firstly,on the mean-teacher model,local and global features are combined,and the attention feature fusion module is introduced to achieve effective fusion,and finally the wheatcount feature maps are processed to generate the predicted density maps and attention maps in a fine-grained way by the multi-scale expansion convolution of the regression head and the attention generator.In addition,a masked image modeling strategy is introduced to facilitate the ability to understand the features.The experimental data show that the semi-supervised model proposed in this paper outperforms other semi-supervised models in the comparison experiments where the proportion of labeled data increases from 10%to 70%.With only 30%labeled data,the model achieves an excellent performance of average absolute error and root mean square error of6.71 and 9.02,respectively,proving its superiority in limited labeled data scenarios.3)Enhancements in deep learning network performance usually come at the cost of increased network size and computational complexity.For counting the number of wheat sheaves in the field,it is also challenging to design lightweight sheaf counting models for edge devices.Therefore,to address this problem,we propose a lightweight wheat ear counting network for efficiently learning global and local features.Specifically,a multi-scale state space group is designed to process multiple blocks of image sequences at different scales,aiming to extract global features at different scales.Next,a lightweight CNN branch is designed to acquire local features.Finally,the global-local adaptive fusion module is used to fuse the multi-scale global and local features by weight dynamic computational mechanism.The experiments show that the lightweight model requires only 7.47M parameters and obtains a mean absolute error of 5.50,a root mean square error of 7.11 and an FPS of 43.更多还原