基于改进YOLOv4模型的橙果识别与定位方法

为提高橙果采摘定位精度和作业速度,提出一种便于迁移至移动终端的改进YOLOv4模型,可从RealSense深度相机所成彩色图像中获取果实质心二维坐标,经配准提取对应深度图中质心点深度值,实现果实的三维空间定位。改进YOLOv4模型以MobileNet v2为主干网络,在颈部结构中使用深度可分离卷积替换普通卷积,实现模型轻量化并提高检测速度。训练后的改进模型对513张独立橙果测试集数据的识别平均精度达97.24%,与原始YOLOv4模型相比,平均检测时间减少11.39 ms,模型大小减少197.5 M。与经典Faster RCNN、SSD模型相比,检测平均精度分别提高了2.69和3.11个百分点,模型大小分别减少了474.5和44.1 M。与轻量化模型YOLOv4-tiny相比,召回率提升了4.22个百分点,较Ghostnet-YOLOv4,平均检测时间减少了7.15 ms。为验证该改进算法实用性,应用改进模型获取果园中78个橙果的位置信息,结果表明：果实二维识别成功率达98.72%,水平方向及垂直方向的平均绝对百分比误差均在1%以内。果实三维定位成功率达96.15%,深度信息平均绝对百分比误差为2.72%,满足采摘机械手精准定位需求。该方法为复杂场景下采摘作业实现提供了鲁棒性强、实时性好、精准度高的目标定位途径。     

基于深度学习的诱捕器内红脂大小蠹检测模型

红脂大小蠹是危害我国北方地区松杉类针叶树种的重大林业入侵害虫,其虫情监测是森林虫害防治的重要环节。传统的人工计数方法已经无法满足现代化红脂大小蠹监测的需求。为自动化识别并统计信息素诱捕器捕获的红脂大小蠹,在传统信息素诱捕器中集成摄像头,自动采集收集杯内图像,建立蠹虫数据集。使用K-means聚类算法优化Faster R-CNN深度学习目标检测模型的默认框,并使用GPU服务器端到端地训练该模型,实现了诱捕器内任意姿态红脂大小蠹的目标检测。采用面向个体的定量评价和面向诱捕器的定性评价两种评价方式。实验结果表明：较原始Faster R-CNN模型,该模型在困难测试集上面向个体和诱捕器的精确率-召回率曲线下面积（Area under the curve,AUC）提升了4.33%和3.28%。在整体测试集上个体和诱捕器AUC分别达0.9350、0.9722。该模型的检测速率为1.6s/幅,准确度优于SSD、Faster R-CNN等目标检测模型,对姿态变化、杂物干扰、酒精蒸发等有较好的鲁棒性。改进后的模型可从被诱芯吸引的6种小蠹科昆虫中区分出危害最大的红脂大小蠹,自动化地统计诱捕器内红脂大小蠹数量。     

高校思想政治理论课教师情怀养成的研究

对高校思想政治理论课教师情怀养成的必要性、有效路径、内容、作用等方面进行了梳理和研究,对进一步丰富关于思想政治理论课教师的相关研究,发挥思想政治理论课教师在教学中的主导作用,提升思想政治理论课教学质量,加强思想政治教育工作队伍建设提供有效借鉴。     

基于深度语义分割的无人机多光谱遥感作物分类方法

为精准获取农田作物种植分布信息以满足农业精细化管理需求,基于DeepLab V3+深度语义分割网络提出了一种面向无人机多光谱遥感影像的农田作物分类方法.通过修改输入层结构、融合多光谱信息和植被指数先验信息、并采用Swish激活函数优化模型,使网络在响应值为负时仍能反向传播.基于2018-2019年连续2年内蒙古自治区河...     

Endoscopic evaluation of the navicular bursa: observations, treatment and outcome in 92 cases with identified pathology

Reasons for performing study: Diagnostic navicular bursoscopy has been described in limited cases. Review of greater numbers is needed to define its contribution to case management and prognostic values. Objectives: To report: 1) clinical, diagnostic and endoscopic findings in a series of cases, 2) surgical techniques and case outcomes and 3) prognostic values. The authors hypothesise: 1) lameness localising to the navicular bursa is commonly associated with dorsal border deep digital flexor tendon (DDFT) lesions, 2) endoscopy allows extent of injuries to be assessed and treated, 3) case outcome relates to severity of DDFT injury and 4) the technique is safe and associated with little morbidity. Materials and methods: All horses that underwent endoscopy of a forelimb navicular bursa for investigation of lameness were identified. Case files were reviewed and those with injuries within the bursa selected for further analysis. Results: One‐hundred‐and‐fourteen horses were identified. Ninety‐two had injuries within the bursa and DDFT injuries were identified in 98% of bursae. Of those examined with magnetic resonance imaging (MRI), 56% had combination injuries involving the DDFT and navicular bone. Sixty‐one percent of horses returned to work sound, 42% returned to previous performance. Horses with extensive tearing and combination injuries of the DDFT and navicular bone identified with MRI, had worse outcomes. Conclusions: Lameness localising to the navicular bursa is commonly associated with injuries to the dorsal border of the DDFT. Endoscopy permits identification and characterisation of injuries within the navicular bursa and enables lesion management. Outcome following debridement is related to severity of injury but overall is reasonable. Potential relevance: Horses with lameness localising to the navicular bursa may have tears of the DDFT. Bursoscopy is able to contribute diagnostic and prognostic information and debridement of lesions improves outcome compared to cases managed conservatively.     

基于YOLO深度卷积神经网络的复杂背景下机器人采摘苹果定位

为提高苹果采摘机器人的工作效率和环境适应性,使其能全天候的在不同光线环境下对遮挡、粘连和套袋等多种情况下的果实进行识别定位,该文提出了基于YOLOv3(you only look once)深度卷积神经网络的苹果定位方法。该方法通过单个卷积神经网络(one-stage)遍历整个图像,回归目标的类别和位置,实现了直接端到端的目标检测,在保证效率与准确率兼顾的情况下实现了复杂环境下苹果的检测。经过训练的模型在验证集下的m AP(meanaverageprecision)为87.71%,准确率为97%,召回率为90%,IOU(intersection over union)为83.61%。通过比较YOLOv3与Faster RCNN算法在不同数目、不同拍摄时间、不同生长阶段、不同光线下对苹果的实际检测效果,并以F1为评估值对比分析了4种算法的差异,试验结果表明YOLOv3在密集苹果的F1高于YOLOv2算法4.45个百分点,在其他环境下高于Faster RCNN将近5个百分点,高于HOG+SVM(histogram of oriented gradient+support vector machine)将近10个百分点。并且在不同硬件环境验证了该算法的可行性,一幅图像在GPU下的检测时间为16.69 ms,在CPU下的检测时间为105.21 ms,实际检测视频的帧率达到了60帧/s和15帧/s。该研究可为机器人快速长时间高效率在复杂环境下识别苹果提供理论基础。     

Combining vision information and tag information to extract Deep Web result pages content

Extracting content from deep web pages is a challenging problem due to the underlying intricate structures of such pages. A vision and tags based approach (DVS) is proposed. It primarily utilizes the vision information and tag information on the Deep Web result pages to extract the content structure of pages. This approach includes two steps as follows: First, the vision information and tag information are produced by analyzing the Cascading Style Sheet and the DOM Tree to generate an initial visual tree of the Deep Web result page. And then, the Path Shingle (PS) algorithm is employed, by considering both of the vision and the tag information, and the blocks in the visual tree are clustered according to the similarity computing result of them to produce the final visual tree, i.e., the content structure of pages. The innovations of DVS are that it utilizes the vision information and tag information on the Deep Web pages to extract the content structure; and stores the vision information as a tree to tansform the analysis of the vision information to a vision attribute tree. Experiments are conducted with a large set of Web databases called UIUC’s TEL. The experimental results show that the vision and tag based approach has high precision compared with the WTS algorithm and the VIPS algorithm.     

基于深度学习的多种农产品供需预测模型

为进一步提高农产品供需过程模拟与估算精度,本研究以自1980年以来国家级和省级的大量农业数据作为样本,充分考虑农产品品种、时间、收入、经济发展等因素影响,构建基于深度学习长短时记忆神经网络（Long Short-Term Memory Neural Network,LSTM）的多种农产品供需预测模型。模型在充分考虑机理性约束条件的前提下,利用深度学习算法在非线性模型分析预测中的优势,对稻谷、小麦、玉米、大豆、猪肉、禽肉、牛肉、羊肉、水产品等9种主要农产品供需进行分析预测。将基于本模型的2019—2021年产量预测结果与国家统计局公布数据进行对比验证,三年平均预测准确率96.98%,表明本研究构建的预测模型能够高效地反映隐性指标变化对预测结果的影响。该模型可以通过及时地监测农业运行数据,为多区域、跨期的农业展望工作提供智能化技术支持。     

华北太行山前平原农田土壤水分动态与氮素的淋溶损失

通过采集土壤溶液并分析其硝态氮（NO3^--N）含量，结合水量平衡方法，研究了华北太行山前平原小麦-玉米轮作农田在当前农民普遍采用的农业管理措施下土壤NO3^--N迁移、累积特征，计算了深层土壤水分渗漏与NO3^--N淋溶损失量。结果表明，土壤水分渗漏、NO3^--N的分布及其淋溶损失存在着明显的时空变异性，土壤水分的深层渗漏和NO3^--N的淋溶损失发生在玉米生长期间施肥灌水或降雨之后。在1998／1999和1999／2000两个作物轮作年中，土壤水分的深层渗漏损失分别为33～48mm（平均39mm）和90～92mm（平均90.7mm），分别占降水＋灌溉总量的10％和19％；淋溶到根区之下的NO3^--N量（包括来自土壤和肥料的N）分别为N12kg hm^-2（范围N6～17kg hm^-2）和N61 kg hm^-2（范围N30～84kg hm^-2），分别占施入肥料量的1．4％～4．1％和7．3％～20．3％。在玉米生长期间有较大潜力可调控灌溉与肥料用量，以提高水肥利用效率。