农业生产环境中的目标识别对象常具有分布密集、体积小、密度大的特点,加之农田环境光照多变、背景复杂,导致已有目标检测模型无法取得令人满意的效果。本研究以提高小目标的识别性能为目标,以蜜蜂识别为例,提出了一种农业复杂环境下尺...农业生产环境中的目标识别对象常具有分布密集、体积小、密度大的特点,加之农田环境光照多变、背景复杂,导致已有目标检测模型无法取得令人满意的效果。本研究以提高小目标的识别性能为目标,以蜜蜂识别为例,提出了一种农业复杂环境下尺度自适应小目标识别算法。算法克服了复杂多变的背景环境的影响及目标体积较小导致的特征提取困难,实现目标尺度无关的小目标识别。首先将原图拆分为一些较小尺寸的子图以提高目标尺度,将已标注的目标分配到拆分后的子图中,形成新的数据集,然后采用迁移学习的方法重新训练并生成新的目标识别模型。在模型的使用中,为使子图识别结果能正常还原,拆分的子图之间需具有一定的重叠率。收集所有子图的目标识别结果,采用非极大抑制(Non-Maximum Suppression,NMS)去除由于模型本身产生的冗余框,提出一种交小比非极大抑制(Intersection over Small NMS,IOS-NMS)进一步去除子图重叠区域中的冗余框。在子图像素尺寸分别为300×300、500×500和700×700,子图重叠率分别为0.2和0.05的情况下进行验证试验,结果表明:采用SSD(Single Shot MultiBox Detector)作为框架中的目标检测模型,新提出的尺度自适应算法的召回率和精度普遍高于SSD模型,最高分别提高了3.8%和2.6%,较原尺度的YOLOv3模型也有一定的提升。为进一步验证算法在复杂背景中小目标识别的优越性,从网上爬取了不同尺度、不同场景的农田复杂环境下的蜜蜂图像,并采用本算法和SSD模型进行了对比测试,结果表明:本算法能提高目标识别性能,具有较强的尺度适应性和泛化性。由于本算法对于单张图像需要多次向前推理,时效性不高,不适用于边缘计算。展开更多
Wireless sensor network(WSN)nodes exchange information via wireless signals,whose power can attenuate at different levels according to the propagation environment.The branches and leaves of young apple trees are much ...Wireless sensor network(WSN)nodes exchange information via wireless signals,whose power can attenuate at different levels according to the propagation environment.The branches and leaves of young apple trees are much sparser than that of adult apple trees.Propagation rules such as propagation distance and attenuation rate are the parameters necessary to know before applying a WSN to a young apple orchard.Field tests were performed,and propagation distance and packet loss rate(PLR)were computed and compared under the two cases:a young apple orchard in fruit period and an open space to find the effect of the apple trees on radio propagation.A model of antenna height and propagation distance was created to forecast the extra path loss caused by the young trees.Validation experiments were performed in a different young apple orchard,and the validation results showed that 70% of R^(2) were higher than 0.7,while the smallest being 0.65;80% RMSE were smaller than 5.The new model was also compared with some classical models such as Cost 235,FITU,ITU-R,and Weissberger model,and the new model was proved to be the best.展开更多
文摘农业生产环境中的目标识别对象常具有分布密集、体积小、密度大的特点,加之农田环境光照多变、背景复杂,导致已有目标检测模型无法取得令人满意的效果。本研究以提高小目标的识别性能为目标,以蜜蜂识别为例,提出了一种农业复杂环境下尺度自适应小目标识别算法。算法克服了复杂多变的背景环境的影响及目标体积较小导致的特征提取困难,实现目标尺度无关的小目标识别。首先将原图拆分为一些较小尺寸的子图以提高目标尺度,将已标注的目标分配到拆分后的子图中,形成新的数据集,然后采用迁移学习的方法重新训练并生成新的目标识别模型。在模型的使用中,为使子图识别结果能正常还原,拆分的子图之间需具有一定的重叠率。收集所有子图的目标识别结果,采用非极大抑制(Non-Maximum Suppression,NMS)去除由于模型本身产生的冗余框,提出一种交小比非极大抑制(Intersection over Small NMS,IOS-NMS)进一步去除子图重叠区域中的冗余框。在子图像素尺寸分别为300×300、500×500和700×700,子图重叠率分别为0.2和0.05的情况下进行验证试验,结果表明:采用SSD(Single Shot MultiBox Detector)作为框架中的目标检测模型,新提出的尺度自适应算法的召回率和精度普遍高于SSD模型,最高分别提高了3.8%和2.6%,较原尺度的YOLOv3模型也有一定的提升。为进一步验证算法在复杂背景中小目标识别的优越性,从网上爬取了不同尺度、不同场景的农田复杂环境下的蜜蜂图像,并采用本算法和SSD模型进行了对比测试,结果表明:本算法能提高目标识别性能,具有较强的尺度适应性和泛化性。由于本算法对于单张图像需要多次向前推理,时效性不高,不适用于边缘计算。
基金financially supported by the National High Technology and Development Program of China(863):Model-based digital management platform for orchard and rape plants(2013AA102405).
文摘Wireless sensor network(WSN)nodes exchange information via wireless signals,whose power can attenuate at different levels according to the propagation environment.The branches and leaves of young apple trees are much sparser than that of adult apple trees.Propagation rules such as propagation distance and attenuation rate are the parameters necessary to know before applying a WSN to a young apple orchard.Field tests were performed,and propagation distance and packet loss rate(PLR)were computed and compared under the two cases:a young apple orchard in fruit period and an open space to find the effect of the apple trees on radio propagation.A model of antenna height and propagation distance was created to forecast the extra path loss caused by the young trees.Validation experiments were performed in a different young apple orchard,and the validation results showed that 70% of R^(2) were higher than 0.7,while the smallest being 0.65;80% RMSE were smaller than 5.The new model was also compared with some classical models such as Cost 235,FITU,ITU-R,and Weissberger model,and the new model was proved to be the best.
