CottonBud-YOLOv5s lightweight cotton bud detection algorithm
-
摘要:
针对棉花机械打顶作业过程中,边缘移动设备算力受限实时性差,运动模糊、小目标遮挡导致难以检测的问题。该研究基于YOLOv5s模型提出CottonBud-YOLOv5s轻量型棉花顶芽检测模型,该模型采用ShuffleNetv2主干网络和DySample动态上采样模块替换原始模块降低计算量,提高模型检测速度;头部(head)和颈部(neck)分别引入ASFFHead检测头和GC(global context)全局注意力模块增强模型尺度不变性和上下文特征提取能力,提高小目标遮挡和运动模糊图像的检测性能。通过消融试验和模型对比试验,验证CottonBud-YOLOv5s棉花顶芽检测模型的可行性。试验结果表明:引入ASFFHead检测头和GC全局注意力机制后,小目标平均精度AP0.5:0.95和平均召回率AR0.5:0.95值分别提升3.6、2.1个百分点,中目标平均精度AP0.5:0.95和平均召回率AR0.5:0.95值分别提升4.1、3.5个百分点,大目标平均精度AP0.5:0.95和平均召回率AR0.5:0.95值分别提升6.5、5.9个百分点;与Faster-RCNN、TOOD、RTDETR、YOLOv3s、YOLOv5s、YOLOv9s和YOLOv10s检测模型相比检测速度分别提升26.4、26.7、24.2、24.8、11.5、18.6、15.6帧,平均精度均值分别提升14.0、13.3、5.5、0.9、0.8、0.2、1.5个百分点,召回率分别提升16.8、16.0、3.2、2.0、0.8、0.5、1.2个百分点,CottonBud-YOLOv5s模型平均精度均值达到97.9%,召回率达到97.2%,CPU检测速度达到27.9帧/s。由模型可视化分析可知CottonBud-YOLOv5s模型在单株、多株、运动模糊、小目标遮挡的整体检测性能优于其他检测模型。该模型具有较高的检测精度、鲁棒性和检测速度,适用于密植环境下棉花顶芽的精准检测,可为棉花机械化打顶提供视觉检测基础。
Abstract:In the context of cotton mechanical topping, several challenges arise due to the limitations of edge-moving devices, including restricted computing power and poor real-time performance. These issues, compounded by phenomena such as motion blur and small target occlusion, significantly hinder the detection process. The focus of this study is to address these challenges by proposing a novel, lightweight cotton bud detection model, named CottonBud-YOLOv5s, which is based on the well-known YOLOv5s architecture. This model incorporates several key improvements to optimize both performance and efficiency in detecting cotton buds in complex field environments. To enhance the model’s overall performance, the CottonBud-YOLOv5s model utilizes the ShuffleNetv2 backbone network, which is specifically chosen for its efficiency in reducing computational complexity while maintaining high detection accuracy. In addition, the DySample dynamic upsampling module is integrated to replace the original upsampling modules, further decreasing computational costs and improving detection speed. These innovations allow the model to run more efficiently on edge devices with limited computing power, addressing the real-time performance issues that often arise during practical applications in cotton mechanical topping. Moreover, the model is designed with an advanced detection head and attention mechanism to bolster its ability to handle varying object scales and complex contextual information. Specifically, the model introduces the ASFFHead detection head and the GC (global context) attention module in the head and neck components, respectively. The integration of these modules enhances the model's scale invariance and significantly improves its capacity for extracting context-based features, which is crucial for detecting small targets that may be occluded or blurred due to motion. These enhancements ultimately improve the model's robustness, enabling it to perform well in challenging real-world conditions. To validate the efficacy of the CottonBud-YOLOv5s model, a series of ablation studies and model comparison tests were conducted. The experimental results demonstrated that the introduction of the ASFFHead detection head and the GC global attention mechanism led to notable improvements in detection accuracy. Specifically, the average precision (AP) at 0.5:0.95 for small targets increased by 3.6 percentage points, while the average recall rate (AR) at the same threshold improved by 2.1 percentage points. For medium-sized targets, the average precision (AP) increased by 4.1 percentage points, and the average recall rate (AR) increased by 3.5 percentage points. For large targets, the average precision (AP) increased by 6.5 percentage points, and the average recall rate (AR) improved by 5.9 percentage points. These results underscore the effectiveness of the proposed enhancements in improving the detection of targets across a range of sizes. Furthermore, when compared to other state-of-the-art detection models, including Faster-RCNN, TOOD, RTDETR, YOLOv3s, YOLOv5s, YOLOv9s, and YOLOv10s, the CottonBud-YOLOv5s model showed significant improvements in detection speed. Specifically, it outperformed these models with speed increases of 26.4, 26.7, 24.2, 24.8, 11.5, 18.6, and 15.6 frames per second, respectively. Additionally, the mean average precision (mAP) was improved by 14.0, 13.3, 5.5, 0.9, 0.8, 0.2, and 1.5 percentage points in comparison to the aforementioned models. The recall rate also saw substantial increases of 16.8, 16.0, 3.2, 2.0, 0.8, 0.5, and 1.2 percentage points, respectively. Overall, the CottonBud-YOLOv5s model achieved a remarkable mean average precision (mAP) of 97.9%, a recall rate of 97.2%, and a CPU detection speed of 27.9 frames per second, demonstrating its exceptional performance in both accuracy and speed. Visual analysis of the model’s performance further confirmed that the CottonBud-YOLOv5s model excels in various detection scenarios, including single-plant, multi-plant, motion blur, and small target occlusion conditions. Its superior performance in these areas highlights its robustness and effectiveness in real-world agricultural environments, where such challenges are commonly encountered. In conclusion, the CottonBud-YOLOv5s model offers a promising solution for precise, real-time detection of cotton buds in densely planted environments. With its high detection accuracy, enhanced robustness, and efficient computational performance, it provides a solid visual detection foundation for cotton mechanized topping, contributing significantly to the advancement of automated agricultural practices.
-
Keywords:
- object detection /
- blocking /
- motion blur /
- small target /
- cotton bud /
- convolutional neural network /
- YOLOv5s
-
0. 引 言
棉花是中国的重要经济作物和战略资源[1]。打顶作业是棉花种植过程中的重要环节,通过去除其主茎顶芽来抑制顶端生长优势,促使棉花多结铃,提高种植效益[2]。目前,棉花打顶主要采用人工打顶、化控打顶和机械打顶3种方式[3]。人工打顶劳动强度大、效率低;化控打顶技术要求高,需要多次喷施,危害劳动人员健康,造成环境污染[4];机械打顶具有节约劳动力、操作简便、效率高等优点,是棉花打顶的必然选择[5]。随着农业科技的发展,机械打顶正朝着智能化和精准化的方向迅速发展[6]。如何提高田间密植环境下棉花顶芽的识别精度,是实现棉花智能化精准打顶作业的关键前提。
早期研究者主要通过激光、红外、超声波等技术检测棉花顶芽位置。史增录等[7]利用激光对射传感器检测植株高度来确定棉花顶芽位置;王玲等[8]运用红外测距实验装置和计算机图像处理技术结合棉花的农学生长特性测量实现了单朵棉花的测距;闫毅敏等[9]设计了一种基于红外测距传感器和超声波测距传感器的棉株顶尖高度测量系统,提高了对棉株顶尖的识别精度。该类方法主要通过检测棉花植株高度来控制机械机构进行打顶,误差较大。
卷积神经网络的出现,推动目标检测进入了全新的阶段,研究人员针对不同的农业应用场景,开展改进算法研究,以进一步提高检测精度和效率[10-13]。范迎迎等[14]提出一种BP神经网络算法的棉花识别方法,能够有效地提高棉花识别精度并简化识别过程;KOIRALA等[15]应用YOLOv3算法来自动识别芒果生长过程中的圆锥花序数量,但检测时实时性较差;张日红等[16]基于YOLOv4检测算法提出轻量化菠萝苗心检测模型,实时性良好但检测准确率有所下降;彭明霞等[17]在RPN网络中引入特征金字塔网络生成目标候选框对Faster R-CNN结构进行优化,在棉间杂草检测上取得较好的效果,但计算量增加检测速度变慢;WU等[18]基于YOLOv4引入残差块增强小目标杂草的检测能力,但模型计算量增大;刘海涛等[19]提出了一种基于YOLOv4算法的识别方法,该方法未能有效识别顶芯,实际打顶效果有待进一步提高。XU等[20]基于YOLOv5检测算法增加P2特征层,增强害虫密集目标检测性能的同时,也造成计算量增大、实时性变差;张新月等[21]提出改进型YOLOv8n轻量化红花识别模型,该模型计算量小、识别速度快,但检测准确率低于基础模型;陈柯屹等[22]提出改进型Faster-RCNN检测算法,解决密植环境中棉花顶芽的精准识别问题,但是Fast R-CNN算法需要采集大量的数据训练模型,在样本不足时检测准确度明显降低;彭炫等[23]基于YOLOv5s算法增加小目标检测层和CPP-CBAM注意力机制,提升了棉花顶芽检测的精度,同时增大了模型计算量。
上述学者针对小目标检测困难、实时性能差等问题开展研究,取得了一些有益效果,但在运动模糊、小目标遮挡检测方面存在巨大挑战。本文以密植环境下棉花顶芽为检测对象,针对棉花打顶作业过程中边缘移动设备算力受限难以满足实时要求,运动模糊、小目标遮挡难以检测的问题,提出一种轻量型棉花顶芽检测模型命名为CottonBud-YOLOv5s,该模型以YOLOv5s模型为基础,通过替换主干网络为ShuffleNetv2轻量级主干网络和DySample轻量化动态上采样模块,显著减少计算量并提升检测速度;引入自适应空间特征融合检测头ASFFHead和GC(global context)全局注意力机制,增强特征的尺度不变性和上下文特征的提取能力,提高对运动模糊图像、小目标遮挡的检测性能,为棉花机械化打顶提供视觉检测基础。
1. 材料与方法
1.1 棉花顶芽图像采集
2023年7月,在山东省滨州市无棣县景国合作社机采棉基地进行棉花顶芽图像的采集工作。采集时间段为每天08:00-19:00。拍摄设备为索尼HDR-CX350E,拍摄方式为垂直向下拍摄棉株顶部。选取棉花顶芽图像1 200张,建立棉花顶芽数据集。数据集包含不同时间段下棉花顶芽单株、多株、运动模糊、遮挡等情况的图像,如图1所示。
1.2 构建数据集
使用LabelImg工具对选取的1 200张原始图片进行单类别标注,标注内容为棉花顶芽,标记为bud。为提升模型泛化能力,通过Python脚本对每张标注好的图片进行数据增强处理,包含上下镜像、左右镜像各一次,随机旋转两次,随机亮度调整一次,合计生成7 200张带有标记的棉花顶芽图像。按照8:1:1的比例将数据增强后的图像随机分为训练集(5 760张)、验证集(720张)和测试集(720张),供后续模型训练和测试任务使用。
2. 棉花顶芽检测算法
2.1 CottonBud-YOLOv5s网络模型
YOLOv5为一阶段目标检测算法,其性能良好,是YOLO算法系列中训练时间和推理速度较优的模型,其中,YOLOv5s网络深度和特征图宽度较小,相较于其他版本推理速度快,更适合部署于边缘移动设备[24-25],本文基于YOLOv5s模型对棉花顶芽检测模型进行改进。
将ShuffleNetv2轻量级主干网络替代原始CSPDarknet53主干网络,同时DySample轻量化动态上采样模块替换最邻近上采样模块,动态调整上采样策略,减少计算量并提高模型检测速度;为提高运动模糊图像、小目标遮挡的检测性能,基于数据驱动金字塔特征融合策略ASFF提出自适应空间特征融合检测头ASFFHead,引入GC全局注意力机制,增强模型尺度不变性和上下文特征提取能力。具体改进方法如下:
第1步,在yolov5s.yaml中更改主干网络为ShuffleNetv2主干网络、更改上采样为DySample动态上采样、更改检测头Head为ASFFHead,添加GC全局注意力机制;第2步,将ShuffleNetv2、DySample、ASFFHead、GC模块增加至common.py文件中;第3步,通过试验进一步调整网络层结构和训练参数,实现改进模块与YOLOv5s目标检测网络的集成,并将改进后模型命名为CottonBud-YOLOv5s,其模型网络结构如图2所示。
2.2 模型轻量化策略
为解决实际应用中边缘移动设备算力受限,模型运行缓慢的问题,基于YOLOv5s模型进行改进优化。采用ShuffleNetv2轻量级主干网络和DySample轻量化动态上采样模块替换原始模块,减少计算量并提高模型检测速度。
2.2.1 ShuffleNetv2轻量级主干网络
ShuffleNetv2[26]是2018年7月由旷视科技提出的轻量级网络,在速度和精度方面进行了均衡。ShuffleNetv2的基本模块如图3a所示,采用通道切分(channel split)将输入特征图在通道维度分成2个通道数相同的分支,左分支为恒等映射减少网络碎片,右分支包含3个通道数一样的卷积使输入输出通道数相同;2个分支进行Concat操作得到特征图的输出通道与原特征图的输入通道相同,其中使用通道混洗(channel shuffle)操作来保证2个分支的信息进行交互。下采样模块如图3b所示,移除通道切分,且2个分支均采用原特征图,并修改卷积单元步长(stride)为2,Concat后特征图大小减半通道数翻倍。
2.2.2 DySample轻量级动态上采样
DySample[27]基于动态采样引入采样点生成器的概念,避免动态卷积的方法,并采用点采样的方式重新定义上采样过程。DySample上采样模块如图4a所示,给定大小为C×H×W的特征图X和大小为2g×H×W的采样集S,其中第一个维度的2代表x和y的坐标值;C为特征图的通道数;H、W为特征图的高度和宽度;g为分组数。
采样函数(grid sample)使用S中的位置将给定的双线性插值X重新采样为大小C×sH×sW的特征图Xs。此过程由式(1)定义:
Xs = grid sample(X,S) (1) 采样集S由采样点发生器(sampling point generator)生成,如图4b所示。DySample上采样引入了“静态偏移范围”和“动态偏移范围”的概念,通过乘以静态因子(0.25)或动态因子(0.5σ)来局部限制采样位置的移动范围,从而减少重叠和预测误差。大小为C×H×W的特征图X通过双线性插值上采样(linear)、静态或动态因子、像素混洗(pixel shuffle)操作将其重塑为大小2g×sH×sW的偏移o,那么采样集S是偏移o和原始采样g的总和,即
o= linear (X) (2) S=g+o (3) 2.3 检测性能增强策略
密植环境下棉花顶芽特征较小易被棉叶遮挡,且棉花打顶作业过程中的运动易导致棉花顶芽图像模糊,给棉花顶芽检测带来困难,为使网络模型能够更准确的检测棉花顶芽位置和特征,对检测头Head改进,并在颈部引入GC全局注意力机制,增强模型尺度不变性和上下文特征提取能力,进而捕捉棉花顶芽的细节特征和增强模型对整体场景的感知能力。
2.3.1 ASFFHead检测头
ASFF[28]自适应特征融合方法主要用于解决Level特征图尺度之间不一致问题,将ASFF嵌入到Head中以增强对不同尺度对象的处理能力。ASFFHead模块结构如图5所示,Level-1、Level-2和Level-3指的是特征金字塔中不同层级的特征,每个层级都有不同的空间分辨率。ASFF-1、ASFF-2和ASFF-3表示应用了ASFF机制的不同层级的特征融合。
ASFF-1是对Level-3的特征图做3×3 MaxPool和3×3 Conv,对Level-2的特征图做3×3 Conv;ASFF-2是对Level-3的特征图做3×3 Conv,对Level-1的特征图做1×1 Conv,并重置为原图分辨率的2倍;ASFF-3是对Level-2的特征图做1×1 Conv,并重置为原图分辨率2倍,对Level-1的特征图做1×1 Conv,并重置为原图分辨率4倍。通过ASFF机制的不同层级的特征融合,所有特征便具有相同的空间维度。
2.3.2 GC注意力机制
视觉领域常用通道注意力机制和空间注意力机制,通道注意力机制使模型能学习不同通道的重要程度,并对其进行加权;其中代表性方法SENet[29]网络,首先对输入特征图进行全局平均池化,获取每个通道的全局信息,再通过两个全连接层和ReLU、Sigmoid激活函数的输出作为通道重要性的权重;将权重与原始输入特征图的各个通道进行加权处理,有效提升重要通道特征的表达能力。空间注意力机制是对空间位置的加权,让模型更加关注目标检测任务相关区域;NLNet(non-local neural networks)网络[30]通过对长距离依赖关系建模,并对远距离信息进行加权融合,增强全局空间信息的关联性。
GC全局注意力机制[31]融合了简化后的Non-local模块和SENet模块的优点,既有NLNet全局上下文建模能力,又像SENet模块一样轻量化。GC全局注意力模块网络结构图如图6所示,首先采用1×1卷积Conv和Softmax函数来获取注意力权重,然后与原始输入特征相乘获得全局上下文特征,上下文信息建模(context modeling)使用NLNet中的机制;特征转换(transform)借鉴SENet模块,GC模块中将C×1×1卷积用瓶颈转换模块代替,该模块将参数量从C×C减少到2×C×C/r,其中r是瓶颈比率。Layer Normalization层的作用是在特征维度上进行归一化,提高模型泛化性。
2.4 试验平台及评价指标
试验训练环境:使用中科视拓科技有限公司的线上服务器(AutoDL),操作系统为 Ubuntu 20.04,内存为120 G,显卡为RTX4090,显存为24 G,CPU为Intel Xeon Gold
6430 16核, CUDA 11.3并行计算框架,Python 3.8.10,PyTorch 1.11.0。试验推理环境:操作系统为 Windows 11,内存为32 G,CPU为Intel i9-12900H 14核,Python 3.8.19,PyTorch 1.13.0。
模型训练参数设置:单机单卡,输入尺寸为3×640×640,每批次样本数量为16,多线程为16,优化器为SGD(stochastic gradient descent),训练次数(epoch)为200轮,初始学习率为0.001,权重衰减系数为
0.0005 ,使用Mosaic、Mixup数据增强方式。为了准确评估模型的性能,采用以下指标[32]:精度P(precision,%),平均精度均值mAP(mean average precision,%),平均精度AP(average precision,%),召回率R(recall,%),参数量(params,M),计算量(flops/G),检测速度FPS(frames per second,帧/s)。其中,精度P表示预测正确的样本数占预测为正样本数的比例,如式(4)所示:
P=TPTP+FP×100% (4) 召回率R表示预测正确的样本数占实际为正样本数的比例,如式(5)所示:
R=TPTP+FN×100% (5) 平均精度AP即为P-R曲线的面积,平均精度均值mAP为平均精确度AP的均值,如式(6)所示:
mAP=N∑i=1∫10P(R)dRN×100% (6) 式中TP表示正确检测棉花顶芽的样本数;FP表示被误测为棉花顶芽的样本数;FN表示未被正确检测的棉花顶芽样本数;N表示类别数量,本文为棉花顶芽,故N=1。
3. 试验结果与分析
3.1 模型改进试验结果
3.1.1 ShuffleNetv2主干网络改进试验
本试验以YOLOv5s目标检测模型为基础,遵循单一变量原则,仅替换不同轻量级主干网络MobileOne、MobileNetV3、GhostNet、ShuffleNetv2进行比较。由表1可知,使用ShuffleNetv2轻量级主干网络作为YOLOv5s主干网络时,计算量为2.20G,精度P为95.4%,平均精度均值mAP为96.2%,召回率R为94.8%,检测速度为40.7帧/s。由于ShuffleNetv2引入的通道切分、通道混洗操作避免分组卷积,增强通道间信息交互,减少计算量并提高了特征提取能力,与改进前CSPDarkNet53主干网络相比,计算量约为原来的1/8,检测速度约为原来的2.5倍,检测精度P、平均精度均值mAP、召回率R分别减小0.7、0.9、1.6个百分点,略有降低。
主干网络
Backbone network计算量Flops/G 检测速度
FPS/(帧·s−1)P/% mAP/% R/% CSPDarkNet53 15.80 16.4 96.1 97.1 96.4 MobileOne 8.00 23.5 88.7 91.9 91.4 MobileNetV3 1.30 34.0 81.8 84.0 81.3 GhostNet 13.40 16.7 95.8 96.7 96.1 ShuffleNetv2 2.20 40.7 95.4 96.2 94.8 注:P为检测精度;mAP为平均精度均值;R为召回率。 Note: P is the detection precision; mAP is the mean average precision; R is the recall rate. GhostNet主干网络的检测速度不满足实时性要求,与MobileOne、MobileNetV3主干网络相比,ShuffleNetv2主干网络检测精度P分别提升6.7、13.6个百分点,平均精度均值mAP分别提升4.3、12.2个百分点,召回率R分别提升3.4、13.5个百分点。因此,本文选择提取特征充分且检测速度都相对良好的ShuffleNetv2主干网络对YOLOv5s模型进行设计与改进。
3.1.2 Dysample上采样模块改进试验
为了弥补替换轻量级主干网络ShuffleNetv2带来的精度损失,同时保持轻量化。使用Dysample上采样模块替换Nearest原始上采样模块,并对比不同上采样模块Nearest、Bilibear、Carafe、Dysample对模型性能指标的影响,试验结果如表2所示。
上采样模块
Upsampling module检测速度
FPS/(帧·s−1)P/% mAP/% R/% Nearest 40.7 95.4 96.2 94.8 Bilinear 41.6 95.1 96.0 95.9 Carafe 38.3 94.7 95.9 95.0 Dysample 40.0 95.5 96.8 95.9 由表2可知,4种上采样模块的检测速度均满足实时性要求,其中,Dysample动态上采样模块与Nearest、Bilinear、Carafe上采样模块相比,检测精度P分别提升0.1、0.4、0.8个百分点,平均精度均值mAP分别提升0.6、0.8、0.9个百分点,与Nearest、Carafe上采样模块相比召回率R分别提升1.1、0.9个百分点,与Bilinear上采样模块召回率持平。替换ShuffleNetv2轻量级主干网络和Dysample动态上采样模块对YOLOv5s进行轻量化改进时,改进后模型在参数量、计算量、检测速度上明显优于改进前的YOLOv5s模型,在检测精度P、平均精度均值mAP、召回率R上也明显优于其他轻量化组合。故选用ShuffleNetv2轻量级主干网络和Dysample动态上采样模块进行轻量化改进是较优组合。
3.1.3 ASFFHead检测头改进试验
替换ShuffleNetv2轻量级主干网络和Dysample动态上采样模块对YOLOv5s进行轻量化改进后,为提高模型对运动模糊图像、小目标遮挡的检测性能,提出ASFFHead改进型检测头。为了测试改进型检测头ASFFHead的性能,与DecoupledHead、CLLAHead、原始Head检测头进行对比试验,试验结果如表3所示。
检测头
Detection head计算量Flops/G P/% mAP/% R/% Head 2.30 95.4 96.2 94.8 CLLAHead 2.90 94.9 96.5 94.6 DecoupledHead 12.30 96.3 97.4 97.0 ASFFHead 4.40 95.5 96.8 95.9 由表3可知,DecoupledHead在精度P、平均精度均值mAP、召回率R上略高于其他检测头,但DecoupledHead与Head、CLLAHead、ASFFHead相比,计算量分别增加10、9.4、7.9G,计算量大量增加导致推理时间过长。ASFFHead与Head、CLLAHead相比,在精度P、平均精度均值mAP、召回率R增加的同时计算量增加较小。ASFFHead检测头通过自适应地调整不同尺度特征的权重,ASFF通过特征金字塔网络(FPN)生成多个尺度的特征图,不同尺度的特征图与对应的权重图逐像素相乘,再进行加权求和,生成融合后的特征图,该特征图在空间和语义上具有高度一致性,从而提高模型对小目标的感知能力,更准确地检测小目标。
3.1.4 GC注意力机制改进试验
在使用ShuffleNetv2轻量级主干网络、Dysample动态上采样模块、ASFFHead检测头改进YOLOv5s的基础上,为测试GC注意力机制对模型的影响,颈部引入CBAM、GE、MLCA、GC不同注意力模块进行对比试验,试验结果如表4所示。
注意力机制
Attention mechanismsP/% mAP/% R/% CBAM 94.2 95.7 94.7 GE 95.2 96.9 95.4 MLCA 95.1 95.3 92.4 GC 96.9 97.9 97.2 由表4可知,GC注意力机制与CBAM、GE、MLCA注意力机制相比,精度P分别提升2.7、1.6、1.8个百分点,平均精度均值mAP分别提升2.2、1.0、2.6个百分点,召回率R分别提升2.5、1.8、4.8个百分点。GC全局注意力机制融合了Non-local模块的优势,继承了NLNet的全局上下文建模能力,保留了更多棉花顶芽的细节特征,Layer Normalization层的引入在特征维度上进行归一化处理,进一步提升模型的泛化能力。
为了更直观地展示GC注意力机制对特征细节的增强作用,采用Grad-CAM技术进行可视化解释和分析,GC注意力机制引入前后,模型对棉花顶芽识别的热力图如图7所示。
由图7可知,引入GC注意力模块前的模型在识别棉花顶芽时,关注部位杂乱且分散,且聚焦部位包含较多背景信息。而引入GC注意力模块后的模型,在单株、多株、运动模糊和遮挡多种情况下,所提模型都将主要注意力聚焦在了棉花顶芽部位。棉花顶芽识别的热力图可视化对比表明,引入GC注意力模块后的模型具备更加高效准确的特征提取能力,且所提模型在识别棉花顶芽时所依据的图像特征信息可信。
3.1.5 CottonBud-YOLOv5s模型消融试验
为了验证模型轻量化后引入ASFFHead检测头和GC全局注意力模块对大、中、小目标检测性能的影响。首先,定义目标框与原图面积的比值小于
0.0025 为小目标,大于0.0225 为大目标,其余为中目标[33]。随后,通过消融试验评估各模块对各类目标检测性能的影响,结果如表5所示。ASFFHead模块
ASFFHead moduleGC注意力模块
GC attention module参数量Params /M 计算量
Flops/GSmall Medium Large AP0.5:0.95/% AR0.5:0.95/% AP0.5:0.95/% AR0.5:0.95/% AP0.5:0.95/% AR0.5:0.95/% − − 4.05 2.30 31.2 41.4 48.2 56.3 61.1 67.0 √ − 9.49 4.40 30.9 37.6 51.4 59.3 64.7 70.5 − √ 4.21 2.30 24.1 31.1 43.1 53.0 57.1 63.6 √ √ 9.65 4.40 34.8 43.5 52.3 59.8 67.6 72.9 注:“−”表示不使用此模块,“√”表示使用此模块;Small、Medium、Large分别代表小目标、中目标、大目标;AP0.5:0.95代表IOU阈值0.5到0.95的平均精度值,AR0.5:0.95代表IOU阈值0.5到0.95的平均召回率。 Note: "−" indicates not using this module, "√" indicates using this module; Small, Medium, Large represent small, medium, and large objects respectively; AP0.5:0.95 represents the average precision value over the IOU threshold range from 0.5 to 0.95, and AR0.5:0.95 represents the average recall value over the IOU threshold range from 0.5 to 0.95. 由表5可知,与不使用ASFFHead检测头和GC注意力机制模块时相比,仅GC全局注意力模块时,大中小目标AP0.5:0.95和AR0.5:0.95值均有所下降;仅使用ASFFHead检测头时,中目标和大目标AP0.5:0.95和AR0.5:0.95值均有所提高,小目标AP0.5:0.95和AR0.5:0.95值均略有下降;同时使用ASFFHead检测头和GC全局注意力模块时,小目标AP0.5:0.95和AR0.5:0.95值分别提升3.6、2.1个百分点,中目标AP0.5:0.95和AR0.5:0.95值分别提升4.1、3.5个百分点,大目标AP0.5:0.95和AR0.5:0.95值分别提升6.5、5.9个百分点,参数量、计算量略有增加。
通过消融试验可知,在棉花顶芽数据集中,轻量化改进后的模型引入ASFFHead检测头和GC注意力模块对大中小目标检测性能均有提升,可满足模型轻量化的同时提高棉花顶芽的检测精度。
3.2 不同目标检测算法对比试验
基于YOLOv5s改进的CottonBud-YOLOv5s模型与目前主流的目标检测模型Faster-RCNN、TOOD、RTDETR、YOLOv3s、YOLOv5s、YOLOv9s、YOLOv10s进行对比试验,结果如表6所示。CottonBud-YOLOv5s模型与Faster-RCNN、TOOD、RTDETR、YOLOv3s、YOLOv5s、YOLOv9s、YOLOv10s相比,计算量分别减小173.6、163.6、126.1、278.6、11.4、23.0、20.4G,检测速度分别提升26.4、26.7、24.2、24.8、11.5、18.6、15.6帧,平均精度均值mAP分别提升14、13.3、5.5、0.9、0.8、0.2、1.5个百分点,召回率R分别提升16.8、16、3.2、2.0、0.8、0.5、1.2个百分点。在棉花顶芽数据集上,CottonBud-YOLOv5s模型在计算量、检测速度、平均精度均值mAP、召回率R上都优于其他7种模型,满足在边缘移动设备上棉花顶芽检测对精度和实时性要求。
模型
Model主干网络
Backbone network计算量
Flops/G检测速度FPS
/(帧·s−1)mAP/% R/% Faster-RCNN Resnet50 178.00 1.5 83.9 80.4 TOOD Resnet101 168.00 1.2 84.6 81.2 RTDETR Resnet50 130.50 3.7 92.4 94.0 YOLOv3s Darknet53 283.00 3.1 97.0 95.2 YOLOv5s CSPDarknet53 15.80 16.4 97.1 96.4 YOLOv9s GELAN 27.40 9.3 97.7 96.7 YOLOv10s CSPNet 24.80 12.3 96.4 96.0 CottonBud-YOLOv5s ShuffleNetv2 4.40 27.9 97.9 97.2 3.3 模型检测效果可视化
为了更加清楚地观察改进效果,针对单目标、多目标、运动模糊、遮挡这4种情况,各随机选取一张照片进行目标检测。对比Faster-RCNN、TOOD、RTDETR、YOLOv3s、YOLOv5s、YOLOv9s、YOLOv10s、CottonBud-YOLOv5s不同模型检测效果,如图8所示。
单株、多株棉花顶芽检测时,Faster-RCNN、TOOD、YOLOv10s模型存在漏检问题;运动模糊图像检测时,Faster-RCNN、TOOD、RTDETR、YOLOv3s、YOLOv5s、YOLOv9s、YOLOv10s模型均存在漏检,仅CottonBud-YOLOv5s正确检测到棉花顶芽;遮挡的棉花顶芽检测时,Faster-RCNN模型存在漏检现象,TOOD、RTDETR、YOLOv3s、YOLOv5s、YOLOv9s、YOLOv10s、CottonBud-YOLOv5s模型均可以正确检测棉花顶芽。CottonBud-YOLOv5s模型在单目标、多目标、运动模糊、遮挡场景下棉花顶芽检测的综合性能最好。
基于以上试验分析,CottonBud-YOLOv5s模型轻量化的同时,提升了运动模糊图像、小目标遮挡的检测性能,可满足棉花顶芽检测实时性与精准性的要求。
4. 结 论
针对边缘移动设备算力受限难以满足实时性要求,棉花打顶作业过程中的运动模糊图像、小目标遮挡难以检测的问题,基于YOLOv5s算法模型,提出了一种轻量化CottonBud-YOLOv5s棉花顶芽检测模型,主要结论如下:
1) 通过将YOLOv5s主干网络和上采样模块分别替换为ShuffleNetv2轻量级主干和Dysample轻量级上采样模块,引入全新自适应空间特征融合检测头ASFFHead 和GC全局注意力模块,有效减少了模型的计算量,提升了检测速度,增强了模型对运动模糊和小目标遮挡的检测性能。
2) CottonBud-YOLOv5s棉花顶芽检测模型平均精度均值mAP达到97.9%,召回率R达到97.2%,CPU检测速度达到27.9帧/s,与主流目标检测模型Faster-RCNN、TOOD、RTDETR、YOLOv3s、YOLOv5s、YOLOv9s和YOLOv10s进行试验比较,检测速度分别提升26.4、26.7、24.2、24.8、11.5、18.6、15.6帧,平均精度均值mAP分别提升14、13.3、5.5、0.9、0.8、0.2、1.5个百分点,召回率R分别提升16.8、16、3.2、2.0、0.8、0.5、1.2个百分点。
3) 可视化结果表明,在单株、多株、运动模糊、小目标遮挡情况下CottonBud-YOLOv5s的检测性能优于其他检测模型,具有较高的检测精度、鲁棒性和检测速度,适用于密植环境下棉花顶芽的精准检测,为棉花机械化打顶提供视觉检测基础。
未来将继续优化模型的泛化能力、扩展其在多样农业场景中的应用、提升在低端硬件上的实时性能、增强模型的鲁棒性,并通过融合多模态数据进一步提升检测精度,发展成为智慧农业领域的关键技术,推动农业现代化的进程。
-
图 2 CottonBud-YOLOv5s模型结构
注:Stem模块为Conv+MaxPool+Concat;ShuffleNetv2为轻量级主干网络;SPPF为空间金字塔池化结构;Conv模块为Conv2 d+BN+SiLU,Conv2 d为二维卷积层,BN表示批量归一化,SiLU为激活函数;C3由Conv模块集成的卷积结构;Dysample为动态上采样模块;Concat为张量拼接操作;GC为Global Context全局注意力模块;MaxPool为最大池化操作;ASFFDetect为自适应空间特征融合检测头。
Figure 2. The CottonBud-YOLOv5s model architecture
Note: The Stem module is Conv+MaxPool+Concat; ShuffleNetv2 is a lightweight backbone network; SPPF is a spatial pyramid pooling structure; the Conv module is Conv2 d+BN+SiLU, Conv2 d is a two-dimensional convolutional layer, BN represents batch normalization, and SiLU is an activation function; C3 is a convolution structure integrated by the Conv module; Dysample is a dynamic upsampling module; Concat is a tensor concatenation operation; GC is a Global Context global attention module; MaxPool is a maximum pooling operation; ASFFDetect is an adaptive spatial feature fusion detection head.
表 1 不同主干网络棉花顶芽试验
Table 1 Test of different backbone networks on cotton buds
主干网络
Backbone network计算量Flops/G 检测速度
FPS/(帧·s−1)P/% mAP/% R/% CSPDarkNet53 15.80 16.4 96.1 97.1 96.4 MobileOne 8.00 23.5 88.7 91.9 91.4 MobileNetV3 1.30 34.0 81.8 84.0 81.3 GhostNet 13.40 16.7 95.8 96.7 96.1 ShuffleNetv2 2.20 40.7 95.4 96.2 94.8 注:P为检测精度;mAP为平均精度均值;R为召回率。 Note: P is the detection precision; mAP is the mean average precision; R is the recall rate. 表 2 不同上采样模块棉花顶芽对比实验
Table 2 Test of different upsampling modules on cotton buds
上采样模块
Upsampling module检测速度
FPS/(帧·s−1)P/% mAP/% R/% Nearest 40.7 95.4 96.2 94.8 Bilinear 41.6 95.1 96.0 95.9 Carafe 38.3 94.7 95.9 95.0 Dysample 40.0 95.5 96.8 95.9 表 3 不同检测头棉花顶芽对比实验
Table 3 Comparison of different detection heads on cotton buds
检测头
Detection head计算量Flops/G P/% mAP/% R/% Head 2.30 95.4 96.2 94.8 CLLAHead 2.90 94.9 96.5 94.6 DecoupledHead 12.30 96.3 97.4 97.0 ASFFHead 4.40 95.5 96.8 95.9 表 4 不同注意力机制对比
Table 4 Comparison of different attention mechanisms
注意力机制
Attention mechanismsP/% mAP/% R/% CBAM 94.2 95.7 94.7 GE 95.2 96.9 95.4 MLCA 95.1 95.3 92.4 GC 96.9 97.9 97.2 表 5 CottonBud-YOLOv5s消融试验
Table 5 CottonBud-YOLOv5s ablation experiments
ASFFHead模块
ASFFHead moduleGC注意力模块
GC attention module参数量Params /M 计算量
Flops/GSmall Medium Large AP0.5:0.95/% AR0.5:0.95/% AP0.5:0.95/% AR0.5:0.95/% AP0.5:0.95/% AR0.5:0.95/% − − 4.05 2.30 31.2 41.4 48.2 56.3 61.1 67.0 √ − 9.49 4.40 30.9 37.6 51.4 59.3 64.7 70.5 − √ 4.21 2.30 24.1 31.1 43.1 53.0 57.1 63.6 √ √ 9.65 4.40 34.8 43.5 52.3 59.8 67.6 72.9 注:“−”表示不使用此模块,“√”表示使用此模块;Small、Medium、Large分别代表小目标、中目标、大目标;AP0.5:0.95代表IOU阈值0.5到0.95的平均精度值,AR0.5:0.95代表IOU阈值0.5到0.95的平均召回率。 Note: "−" indicates not using this module, "√" indicates using this module; Small, Medium, Large represent small, medium, and large objects respectively; AP0.5:0.95 represents the average precision value over the IOU threshold range from 0.5 to 0.95, and AR0.5:0.95 represents the average recall value over the IOU threshold range from 0.5 to 0.95. 表 6 不同目标检测模型对比
Table 6 Comparison of different object detection models
模型
Model主干网络
Backbone network计算量
Flops/G检测速度FPS
/(帧·s−1)mAP/% R/% Faster-RCNN Resnet50 178.00 1.5 83.9 80.4 TOOD Resnet101 168.00 1.2 84.6 81.2 RTDETR Resnet50 130.50 3.7 92.4 94.0 YOLOv3s Darknet53 283.00 3.1 97.0 95.2 YOLOv5s CSPDarknet53 15.80 16.4 97.1 96.4 YOLOv9s GELAN 27.40 9.3 97.7 96.7 YOLOv10s CSPNet 24.80 12.3 96.4 96.0 CottonBud-YOLOv5s ShuffleNetv2 4.40 27.9 97.9 97.2 -
[1] 李继福,何俊峰,陈佛文,等. 中国棉花生产格局与施肥研究现状——基于CNKI数据计量分析[J]. 中国棉花,2019,46(4):17-24. doi: 10.11963/1000-632X.ljfljf.20190402 LI Jifu, HE Junfeng, CHEN Fowen, et al. Research on the production pattern and fertilization of cotton in china: Based on CNKI data metric analysis[J]. Chinese Cotton, 2019, 46(4): 17-24. (in Chinese with English abstract) doi: 10.11963/1000-632X.ljfljf.20190402
[2] 韩长杰,燕长辉,仇世龙,等. 基于机器视觉的双圆盘式棉花打顶装置设计与试验[J]. 农业机械学报,2023,54(5):36-46. doi: 10.6041/j.issn.1000-1298.2023.05.004 HAN Changjie, YAN Changhai, QIU Shilong, et al. Design and experiment of double disc cotton topping device based on machine vision[J]. Transactions of the Chinese Society for Agricultural Machinery, 2023, 54(5): 36-46. (in Chinese with English abstract) doi: 10.6041/j.issn.1000-1298.2023.05.004
[3] 李涛,杨德松. 棉花不同打顶方式应用分析[J]. 新疆农垦科技,2021,44(2):14-15. doi: 10.3969/j.issn.1001-361X.2021.02.008 LI Tao, YANG Desong. Analysis of different topping methods for cotton[J]. Xinjiang Agricultural Science and Technology, 2021, 44(2): 14-15. (in Chinese with English abstract) doi: 10.3969/j.issn.1001-361X.2021.02.008
[4] 韩鑫,韩金鸽,陈允琳,等. 基于顶芽智能识别的棉花化学打顶系统研究[J]. 农业机械学报,2024,55(3):145-152. doi: 10.6041/j.issn.1000-1298.2024.03.014 HAN Xin, HAN Jinge, CHEN Yunlin, et al. Research on the chemical topping system of cotton based on apical bud intelligent recognition[J]. Transactions of the Chinese Society for Agricultural Machinery, 2024, 55(3): 145-152. (in Chinese with English abstract) doi: 10.6041/j.issn.1000-1298.2024.03.014
[5] 彭强吉,康建明,宋裕民,等. 3MDZ-4型自走式棉花打顶喷药联合作业机设计[J]. 农业工程学报,2019,35(14):30-38. doi: 10.11975/j.issn.1002-6819.2019.14.004 PENG Qiangji, KANG Jianming, SONG Yumin, et al. Design of 3MDZ-4 self-propelled cotton topping and spraying joint operation machine[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2019, 35(14): 30-38. (in Chinese with English abstract) doi: 10.11975/j.issn.1002-6819.2019.14.004
[6] 卢秀茹,贾肖月,牛佳慧. 中国棉花产业发展现状及展望[J]. 中国农业科学,2018,51(1):26-36. doi: 10.3864/j.issn.0578-1752.2018.01.003 LU Xiuru, JIA Xiaoyue, NIU Jiahui. Current situation and prospect of cotton industry development in China[J]. Chinese Agricultural Science, 2018, 51(1): 26-36. (in Chinese with English abstract) doi: 10.3864/j.issn.0578-1752.2018.01.003
[7] 史增录,张学军,丁永前,等. 基于自动检测与控制的棉花打顶装置的设计[J]. 中国农机化学报,2017,38(12):58-62. SHI Zenglu, ZHANG Xuejun, DING Yongqian, et al. Design of cotton topping device based on automatic detection and control[J]. Transactions of the Chinese Society of Agricultural Machinery, 2017, 38(12): 58-62. (in Chinese with English abstract)
[8] 王玲,邹小昱,刘思瑶,等. 棉花采摘机器人红外测距技术研究[J]. 农业机械学报,2014,45(7):61-66. WANG Ling, ZOU Xiaoyu , LIU Siyao, et al. Infrared distance measurement used for cotton picker robot[J]. Transactions of the Chinese Society for Agricultural Machinery, 2014, 45 (7): 61-66. (in Chinese with English abstract)
[9] 闫毅敏,王维新,李霞,等. 基于自动控制的棉株顶尖高度测量系统设计与试验[J]. 甘肃农业大学学报,2018,53(5):176-184. YAN Yimin, WANG Weixin, LI Xia. Design and test of cotton top height measurement system based on automatic control[J]. Journal of Gansu Agricultural University, 2018, 53(5): 176-184. (in Chinese with English abstract)
[10] GANESAN G, CHINNAPPAN J. Hybridization of ResNet with YOLO classifier for automated paddy leaf disease recognition: An optimized model[J]. Journal of Field Robotics, 2022, 39(7): 1085-1109. doi: 10.1002/rob.22089
[11] LI D, AHMED F, WU N, et al. Yolo-JD: A Deep Learning Network for jute diseases and pests detection from images[J]. Plants, 2022, 11(7): 937. doi: 10.3390/plants11070937
[12] BAKHSHAYESHI I, ERFANI E, TAGHIKHAH F R, et al. An intelligence cattle re-identification system over transport by siamese neural networks and YOLO[J]. IEEE Internet of Things Journal, 2023, 11(2): 2351-2363.
[13] PEREZ-PORRAS F J, TORRES-SANCHEZ J, LOPEZ-GRANADOS F, et al. Early and on-ground image-based detection of poppy (Papaver rhoeas) in wheat using YOLO architectures[J]. Weed Science, 2023, 71(1): 50-58. doi: 10.1017/wsc.2022.64
[14] 范迎迎,钱育蓉,杨柳,等. 基于BP神经网络的遥感影像棉花识别方法[J]. 计算机工程与设计,2017,38(5):1356-1360. FANG Yingying, QIAN Yurong, YANG Liu, et al. Cotton recognition method for remote sensing image based on BP neural network[J], Computer Engineering and Design, 2017, 38(5): 1356-1360. (in Chinese with English abstract)
[15] KOIRALA A, WALSH K B, WANG Z, et al. Deep learning for mango (Mangifera indica) panicle stage classification[J]. Agronomy, 2020, 10(1): 2073-4395.
[16] 张日红,区建爽,李小敏,等. 基于改进YOLOv4的轻量化菠萝苗心检测算法[J]. 农业工程学报,2023,39(4):135-143. ZHANG Rihong, QU Jianshuang, LI Xiaomin, et al. Lightweight pineapple seedling heart detection algorithm based on improved YOLOv4[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2023, 39(4): 135-143. (in Chinese with English abstract)
[17] 彭明霞,夏俊芳,彭辉. 融合FPN的Faster R-CNN复杂背景下棉田杂草高效识别方法[J]. 农业工程学报,2019,35(20):202-209. doi: 10.11975/j.issn.1002-6819.2019.20.025 PENG Mingxia, XIA Junfang, PENG Hui. Efficient weed identification method in cotton field with complex background based on Faster R-CNN fused with FPN[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2019, 35(20): 202-209. (in Chinese with English abstract) doi: 10.11975/j.issn.1002-6819.2019.20.025
[18] WU H, WANG Y, ZHAO P, et al. Small-target weed-detection model based on YOLO-V4 with improved backbone and neck structures[J]. Precision Agriculture, 2023, 24(6): 2149-2170. doi: 10.1007/s11119-023-10035-7
[19] 刘海涛,韩鑫,兰玉彬,等. 基于YOLOv4网络的棉花顶芽精准识别方法[J]. 中国农业科技导报,2022,24(8):99-108. LIU Haitao, Han Xin, LAN Yubin, et al. Precision identification of cotton terminal bud based on YOLOv4 network[J]. Journal of Agricultural Science and Technology ofChina, 2022, 24(8): 99-108. (in Chinese with English abstract)
[20] XU W, XU T, THOMASSON J A, et al. A lightweight SSV2-YOLO based model for detection of sugarcane aphids in unstructured natural environments[J]. Computers and Electronics in Agriculture, 2023, 211: 107961. doi: 10.1016/j.compag.2023.107961
[21] 张新月,胡广锐,李浦航,等. 基于改进YOLOv8n的轻量化红花识别方法[J]. 农业工程学报,2024,40(13):163-170. doi: 10.11975/j.issn.1002-6819.202403036 ZHANG Xinyue, HU Guangrui, LI Puhang, et al. Recognizing safflower using improved lightweight YOLOv8n[J]. Journal of Agricultural Engineering, 2024, 40(13): 163-170. (in Chinese with English abstract) doi: 10.11975/j.issn.1002-6819.202403036
[22] 陈柯屹,朱龙付,宋鹏,等. 融合动态机制的改进型Faster R-CNN识别田间棉花顶芽[J]. 农业工程学报,2021,37(16):161-168. doi: 10.11975/j.issn.1002-6819.2021.16.020 CHEN Keyi, ZHU Longfu, SONG Peng, et al. Improved Faster R-CNN with fusion dynamic mechanism for identifying cotton apical buds in the field[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2021, 37(16): 161-168. (in Chinese with English abstract) doi: 10.11975/j.issn.1002-6819.2021.16.020
[23] 彭炫,周建平,许燕,等. 改进YOLOv5识别复杂环境下棉花顶芽[J]. 农业工程学报,2023,39(16):191-197. doi: 10.11975/j.issn.1002-6819.202305026 PENG Xuan, ZHOU Jianping, XU Yan, et al. Identifying cotton apical buds in complex environments using improved YOLOv5[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2023, 39(16): 191-197. (in Chinese with English abstract) doi: 10.11975/j.issn.1002-6819.202305026
[24] 白强,高荣华,赵春江,等. 基于改进YOLOV5s网络的奶牛多尺度行为识别方法[J]. 农业工程学报,2022,38(12):163-172. doi: 10.11975/j.issn.1002-6819.2022.12.019 BAI Qiang, GAO Ronghua, ZHAO Chunjiang, et al. Multi-scale behavior recognition of dairy cows based on improved YOLOV5s network[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2022, 38(12): 163-172. (in Chinese with English abstract) doi: 10.11975/j.issn.1002-6819.2022.12.019
[25] 尚钰莹,张倩如,宋怀波. 基于YOLOv5s的深度学习在自然场景苹果花朵检测中的应用[J]. 农业工程学报,2022,38(9):222-229. doi: 10.11975/j.issn.1002-6819.2022.09.024 SHANG Yuying, ZHANG Qianru, SONG Huaibo. Application of deep learning based on YOLOv5s in detecting apple flowers in natural scenes[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2022, 38(9): 222-229. (in Chinese with English abstract) doi: 10.11975/j.issn.1002-6819.2022.09.024
[26] MA N, ZHANG X, ZHENG H T, et al. Shufflenetv2: Practical guidelines for efficient cnn architecture design[C]//Proceedings of the European Conference on Computer Vision (ECCV). Cham, Springer, 2018: 122-138.
[27] LIU W, LU H, FU H, et al. Learning to upsample by learning to sample[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). Paris, USA: IEEE , 2023: 6004-6014.
[28] LIU S, HUANG D, WANG Y. Learning spatial fusion for single-shot object detection DB/OL. (2019-11-25) 2025−01−06. https://doi.org/10.48550/arXiv.1911.09516.
[29] XU G, GAO Y, LI J, et al. InSAR phase denoising: a review of current technologies and future directions[J]. IEEE Geoscience and Remote Sensing Magazine, 2020, 8(2): 64-82. doi: 10.1109/MGRS.2019.2955120
[30] WANG X, GIRSHICK R, GUPTA A, et al. Non-local neural networks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City, USA: IEEE, 2018: 7794-7803.
[31] CAO Y, XU J, LIN S, et al. GCNet: Non-Local Networks meet squeeze-excitation networks and beyond[C]//2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). Seoul , USA: IEEE, 2020: 1971-1980.
[32] 兰玉彬,孙斌书,张乐春,等. 基于改进YOLOv5s的自然场景下生姜叶片病虫害识别[J]. 农业工程学报,2024,40(1):210-216. LAN Yubin, SUN Binshu, ZHANG Lechun, et al. Identification of ginger leaf diseases and pests in natural scenes based on improved YOLOv5s[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2024, 40(1): 210-216. (in Chinese with English abstract)
[33] 顾洪宇,李志合,李涛,等. 基于YOLOv5的马铃薯种薯芽眼轻量化检测算法[J]. 农业工程学报,2024,40(14):126-136. doi: 10.11975/j.issn.1002-6819.202404033 GU Hongyu, LI Zhihe LI Tao, et al. Lightweight detection algorithm for potato seed tubers' bud eyes based on YOLOv5[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2024, 40(14): 126-136. (in Chinese with English abstract) doi: 10.11975/j.issn.1002-6819.202404033