基于多姿态模板匹配的果树性诱靶标害虫识别模型与计数方法研究

Automatic identification and counting for fruit pests real-time monitoring was an important foundation for precise pest-control. Currently, pest image identification researches focus on feature extraction based on standard pose samples, without considering the influence from multi-pose of pests in natural scene. Consequently, the identification rate of these methods is low in practical orchand application. Dichocrocis Punctiferalis was selected for this study, a new method for fruit target pests captured by pheromone-baited counting was put forward based on multi-pose templates matching identification technology to overcome the shortcoming mentioned above. The impact factors of and threshod of the target pests pose sptial distribution were explored, and a novel manifold learning method was adopted to explore the low-dimensional feature space distribution structure hidden in high-dimensional observable pest image space,thus a fruit pest pose sets were established.The global features(color, shape) and affine invariant local features(angular points, interest points) were extracted from different pose images. Then these different information were selected and compressed to establish the multi-pose template library for fruit pest identification. K fold cross-validation method was used to select the template sets and test sets, and classifier parameters were structured and optimized through training, thus a pest identification model was established based on template matching for identification and counting of orchard pests to improve the identification accuracy and robustness and provide the guidance for developing of practical devices of pests counting.

果树害虫自动监测识别与计数是精确防治得以实施的重要基础。目前害虫图像识别集中在害虫样本标准姿态模板下的特征提取，较少考虑自然场景下性诱害虫多姿态对识别的影响，导致应用中识别率偏低。本研究拟以果树常发害虫桃柱螟性诱成虫为例，探究基于多姿态模板匹配的果树性诱靶标害虫识别模型与计数新方法。探究靶标害虫姿态空间分布的影响因子和阈值判定方法，研究基于非线性流行学习的图像空间与姿态空间映射方法，建立靶标害虫姿态集；研究不同姿态下害虫图像特征，挖掘差异信息并进行选择与压缩，筛选图像中颜色、形态等全局特征及角点、兴趣点等尺度不变局部特征指标，建立基于姿态集的害虫识别模板库。采用K折交叉验证法对模板集和测试集进行选择，通过训练、构造和优化分类器参数，建立害虫模板匹配识别模型，构建适于自然场景下果树诱捕害虫识别与计数新方法，达到提高果树害虫识别计数准确率和鲁棒性的目的，为开发实用的害虫计数设备提供方法指导。

果园害虫种类及数量的监测是果园害虫综合治理的基础，目前害虫监测方式主要是依靠人工现场调查与统计完成，费时费力。论文研究了基于机器视觉的果园性诱害虫在线识别与计数方法，探讨了在线监测设备与系统的样机构建，为推进基于图像的害虫自动识别与分类技术在实际果园害虫监测中的应用提供技术支撑。重点开展了以下四方面工作：.（1）提出了一种基于形状因子和分割点定位的害虫图像分割方法。针对实验室环境下的桃蛀螟(Conogethes punctiferalis (Guenée))害虫图像，本文方法平均错误率为7%，约为分水岭分割方法的1/2，平均分割有效率为92.65%，比分水岭算法提高了5.7%；在对田间梨小食心虫(Grapholitha molesta (Busck) )图像分割中，本文方法平均错误率为2.24%，平均分割有效率为97.8%，分别比分水岭方法降低了4.29%和提高了3.95%。.（2）从形状不变特征与形状无关特征两种思路出发，分别研究了基于图像Zernike矩和颜色纹理组合的果园靶标害虫特征提取方法。在害虫图像Zernike矩特征提取中，构建了低阶Zernike矩和高阶Zernike矩两个特征组合，用于验证图像Zernike矩的阶数对果园害虫识别准确率的影响；在基于颜色纹理特征提取中，构建了RGB和HSV两种颜色空间下6种特征组合。.（3）研究了基于多类支持向量机的果园害虫图像识别分类方法。通过利用网格搜索、遗传算法和粒子群优化算法三种方法对支持向量机分类器参数进行寻优得到：基于遗传算法的参数寻优结果表现最优。在此基础上，选择基于DAG-SVM的方法构建本文的多类SVM分类器，利用测试样本，对两类特征的测试识别结果对比表明：基于’db4’小波分解的HSV三通道纹理特征组合性能最优，不仅识别率达到了100%，而且时间成本最低。.（4）果园害虫在线监测设备构建与算法验证。依据实际应用场景及参数对害虫监测设备的硬件和软件进行初步设计与实现。利用构建的监测设备进行了多组害虫识别试验，系统对单种非粘连、轻度粘连的害虫具有较高的识别率，平均识别率在95%以上，而对重度粘连的害虫，平均识别率为84.2%；在多种害虫识别时，分割方法的阈值需要根据害虫体积差异大小进行选择，多种害虫样本混合识别试验的平均识别准确率为87.67%。