基于图谱特征分析的农业虫害检测方法研究

发布时间：2018-03-26 03:19

本文选题：虫害检测　切入点：光谱检测技术　出处：《浙江大学》2017年博士论文

【摘要】：虫害是农业灾害频繁爆发的主要成因之一,在早期发现并准确定位和识别害虫,对其未来的发展趋势做出评估,可提高施药决策和综合防治的针对性和准确性。在作物虫害信息的获取中,传统的检测方法不但耗时、费力,而且滞后性较强,很难满足现代农业的精准生产要求。因此,构建高效、实时的自动化虫害检测方法来替代传统手段,将有助于改变这一现状。本文依托光谱技术和计算机视觉技术作为研究方法,从光谱和图像数据特征分析的角度出发,研究了农业虫害的自动化检测方法,具体开展了以下4方面工作:(1)害虫幼虫隐蔽性强,直接检测较为困难,故采用高光谱成像技术对健康和受幼虫侵染的作物进行了检测研究。针对光谱数据维度高、冗余性强的特点,提出一种栈化稀疏自编码模型(StackedSparseAuto-Encoder,SSAE),实现了对不同虫害侵染阶段间稀疏性判别信息的高效表达。结合SSAE提出了一种基于Softmax函数的自适应损失敏感算法,实现了对早期虫害侵染的敏感性检测。定义了一种相对离散值指标(Relative Scatter Value,RSV),用于量化采用自适应损失敏感算法前后光谱数据的类间和类内分布差异,以直观的方式验证了自适应损失敏感算法为数据表达所引入的线性可分特性。(2)采用计算机视觉技术对农田内害虫成虫的图像定位进行了研究,提出了一种结合图像显著性分析和GrabCut分割的害虫目标自动定位方法。利用害虫目标区域和农田背景在图像全局层面的颜色和空间属性差异,构建了指示害虫潜在区域的显著性映射图(Saliency Map),并通过阈值化Saliency Map提供GrabCut分割算法的初始区域,克服了该算法要求人工干预的缺点,实现了害虫目标的自动定位。从定位准确性和执行时间层面对算法进行优化,将定位准确率提升至0.9以上,单幅图像的平均处理时间维持在150 ms以内。(3)结合卷积神经网络(Convolutional Neural Network,CNN)对害虫成虫图像的视觉分类进行了研究,并基于CNN重要的结构参数和训练策略做出一定的优化和改进。针对害虫图像目标所存在的种间相似性高、种内变化性大、姿态变化丰富、易受作物遮挡等特点,从CNN的底层特征提取、中层特征规模、高层信息组织、分类器选择以及模型训练技术方面,确定了适合农业害虫图像识别的最优模型参数。优化后模型在包含14类稻田害虫的图像数据集上获得0.883的识别准确率,相对传统识别方法效果提升显著。优化后模型内存需求和运行耗时分别降低至6.0 MB和0.7 ms,极大地加强了 CNN在害虫检测任务中的实际可应用性。(4)针对害虫图像采集难度大、标注成本高而导致的数据集规模小、类别分布不均匀问题,研究了结合特征迁移用于提升视觉识别模型性能的方法。以CNN作为基准研究框架,将由大规模广义数据集预训练而来的图像特征迁移至特定害虫图像数据集构建模型初始化参数,实现了相对常规随机初始化方法的显著性能提升。从模型架构层面,对比分析了在随机初始化和特征迁移条件下,不同CNN模型在害虫图像识别中的性能差异。设计CNN特征逐层迁移后参数固定/可调的对照实验,提出了底层特征可直接迁移,中间连续层特征不能单独迁移,高层特征不适合迁移的策略。通过随机改变训练数据集大小和分布特性,验证了特征迁移对于数据规模和类别分布不均匀的相对不敏感特性。通过可视化模型训练/测试损失曲线、底层特征、高层表达以及表达后数据的分布特性,对比并阐明了特征迁移相对随机初始化方法优势明显的主要原因。
[Abstract]:Insect pest is one of the main causes of agricultural disasters broke out frequently, in early detection and accurate positioning and identification of pests, assess its future development trend, can improve the application decision and comprehensive prevention and control of pertinence and accuracy. In the acquisition of crop pest information, the traditional detection methods are time-consuming, laborious, but also behind the strong it is difficult to meet the requirements of production, the precision of modern agriculture. Therefore, constructing an efficient, automated pest real time detection method to replace the traditional means, will help to change this situation. On the basis of spectral technology and computer vision technology as research methods, from the analysis of spectral image data and feature point of view, research on automation of detection the specific method of agricultural pests, carried out the following 4 aspects: (1) insect larvae concealment, direct detection is difficult, so the use of hyperspectral imaging technology for health and The larvae infected crops were measured. According to the spectral data of high dimension, high redundant characteristics, proposed a stack of sparse self encoding model (StackedSparseAuto-Encoder, SSAE), the high expression of discriminant information of different pest infection stage. Among the sparse combination of SSAE presents an adaptive loss sensitive algorithm Softmax based on the function, the sensitivity of detection for early pest infection. The definition of a relatively discrete value index (Relative Scatter Value, RSV), the adaptive algorithm is used to quantify the loss of sensitive data before and after the light spectrum between class and within class distribution, an intuitive way to verify the adaptive algorithm for the linear expression of loss sensitive data the characteristics can be divided. (2) using computer vision technology image positioning of the farmland pests were studied, proposes a combination of image saliency Automatic positioning method and GrabCut analysis of pest target segmentation. The use of pest target area and farmland in the global level background difference image color and spatial attributes, construct the significant map indicating potential PEST region (Saliency Map), and Saliency Map provided by thresholding initial region GrabCut segmentation algorithm, the algorithm overcomes the requirements the shortcomings of manual intervention, realizes the automatic positioning of the target. The pest algorithm is optimized from the positioning accuracy and execution time in positioning accuracy rate increased to more than 0.9, average single image processing time maintained at less than 150 ms. (3) combined with convolutional neural network (Convolutional Neural Network, CNN) on visual classification adult pest images are studied, and based on CNN important structural parameters and training strategies to optimize and improve the existing target pest images. The species similarity is high in change of attitude, variety, characteristics of crops susceptible to occlusion, extracted from the underlying features of CNN, characteristics of middle scale, high-level information organization, classifier selection and model training techniques, to determine the optimal parameters for image recognition model of agricultural pests. The optimized model in image the data contains 14 kinds of insect pests in paddy fields set achieves 0.883 recognition accuracy, compared with the traditional identification methods to enhance the effect significantly. The optimized model memory requirement and operation time is reduced to 6 MB and 0.7 MS respectively, greatly strengthened the actual CNN in pest detection task can be applied. (4) on the basis of pest image acquisition the difficulty, marking the result of the high cost of the data set is small, the problem of the uneven distribution of categories, the research combined with the feature of migration method for improving performance of vision recognition model. Using CNN as the base of the frame Frame model initialization parameters by generalized large-scale data sets and the image features of pre training migration to specific pest image data sets, to achieve significant performance relative to the conventional random initialization method to enhance the model. From the architecture level, a comparative analysis on the random initialization and migration characteristics under the condition of different performance characteristics of CNN model in pest image the design feature recognition. CNN layer after migration parameters fixed / adjustable control experiments, the underlying characteristics of direct migration, middle layer continuous features alone can not transfer, not suitable for the high-level feature migration strategy. By randomly changing the training data set size and distribution characteristics for the transfer of relatively sensitive data the size and type of uneven distribution is verified. Through visualization model training / test loss curve, low-level features, high-level expression and expression of data points The main reasons for the obvious advantages of the relative random initialization method of feature migration are compared and illustrated.

【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：S433

【相似文献】