农林检疫害虫图像自动识别研究

With the global climate becoming more abnormal and agriculture product imports and exports increasing, the loss by harmful insects and foreign species intrusion is more and more severe. So it is necessary for our food and zoology safety to eatabilish large scale inspecting sysytem by automated recognizing harmful farm and quarantine insects by computer. This project is faced on harmful farm and quarantine insect images captured under free pose in natural scene, aims at the small training set, large category set, pose and view point robust feature extraction problems. We will combining computer vision, machine learning and insect appearance characteristic, break through several key technologies and provide technique deposite for large scale application of harmful farm and quarantine insects. Regarding the small training set problem, we will increasing the training samples by transfer learning. Regarding the large large category set problem ,we will design classification techniques for handreds upto thousand categories by nonlinear metric learning and image retrival. Regarding robust feature extraction problem, we start from the shape and appearance characteristic of insects, design pose and view point invariant features by sailent feature extraction, multiple feature fusion and graph matching. We will also provied a specimen dataset for insect recognition research composed of 3 ordos, 17 genus and 100 species including over 2000 insect specimens and corresponding image dataset.

随着全球气候异常和农产品进出口频繁，我国农林病虫害和外来物种入侵造成的损失逐年加重。利用计算机自动识别农林检疫害虫，建立大规模监测体系，事关我国粮食安全和生态安全。本项目面向自然场景中的农林检疫害虫图像，以解决害虫识别中的小训练样本、大类别集、姿态和视角变化的鲁棒特征提取等问题为目标，以计算机视觉、机器学习方法与昆虫图像特点相结合为手段，力求在这几个方面突破关键技术，为自然场景中的农林检疫害虫图像识别应用提供技术储备。针对小训练样本问题，我们将采用迁移学习方法增加训练样本。针对大类别集问题，我们将采用非线性距离度量、图像检索方法设计出适用于百类乃至千类的图像分类方法。在鲁棒特征方面，我们结合昆虫形态特点，利用显著性特征、多特征融合和图匹配等方法设计出姿态和视角不变的鲁棒特征。本项目也涉及大量昆虫标本获取和图像采集，将建立一个包含3个目17个属100个种的2000个昆虫标本以及相关的图像库。

本项目面向农林检疫害虫图像识别，以满足自动化规模化害虫监测替代人类昆虫鉴定专家为需求，以计算机视觉、机器学习结合昆虫图像形态特征为手段，主要研究害虫识别中常见的小样本、大类别集、姿态和视角变化条件下的鲁棒特征提取和识别问题。本项目建立了一个包含4个目9个科34个属100个种的2000多个昆虫标本库以及相应的图像库，基于词袋模型提出了鉴别性词典学习、鉴别性编码、空间位置信息嵌入和软划分汇聚等若干算法，基于Gabor曲面特征结合线性判别分析实现了大类别集昆虫图像分类。针对标本采集过程中的昆虫脆弱部件损毁问题，本项目提出了异质类别集融合的方法，可以处理分类器融合中某些部件和类别缺失的问题，具有一定理论意义。在上述昆虫图像数据库和提出的算法基础上，本项目开发了实蝇科昆虫图像自动识别系统(AFIS1.0)，第一候选识别率在70%以上，前五候选识别率在90%以上，进一步结合昆虫形态特征执行人工判读，识别率达到了95%，并在云南出入境检验检疫局及其下属的三个口岸进行了试用，为口岸有害实蝇的鉴定和检疫奠定了基础，具有一定的实际价值。

内容获取失败，请点击重试