CORC  > 自动化研究所  > 中国科学院自动化研究所  > 毕业生  > 博士学位论文
题名基于视觉的昆虫物种识别算法研究
作者陆安
学位类别工学博士
答辩日期2012-12-05
授予单位中国科学院大学
授予地点中国科学院自动化研究所
导师刘成林 ; 侯新文
关键词昆虫识别 特征融合 判别性编码 特征汇聚 分类器融合 insect recognition feature fusion discriminative coding feature pooling classifier fusion
其他题名Research on Vision-based Insect Species Recognition Algorithm
学位专业模式识别与智能系统
中文摘要昆虫识别技术在动植物检疫、农林业病虫灾害检测与防治、生态系统研究与保护等领域具有重要的应用前景。目前,对昆虫识别技术人员的需求与供给之间形成了巨大缺口,成为了制约我国农业、林业、生态、环保等事业发展的重大瓶颈。随着计算机科学技术的快速发展,计算机的运算处理能力、数字图像处理技术、计算机视觉技术、机器学习与模式识别的理论方法都有了突破性的进展。利用数字图像处理技术和模式识别方法进行昆虫自动识别,从根本上可以解决昆虫学家人员不足的现状,为我国的农林、生态、环保等事业创造巨大的经济价值。 本文针对昆虫图像的特点,在昆虫图像的特征提取、编码方法、汇聚方法、分类器融合等方面进行了深入研究。通过计算机视觉和模式识别技术对昆虫图像进行处理,实现昆虫物种的自动识别,具有理论意义和应用价值。论文主要研究成果如下: 1. 提出一种针对昆虫图像识别的特征层融合方法。充分利用图像的局部纹理、颜色、形状等特征的互补性,将三类局部特征融合后,再进行字典学习、编码和分类,这种方法有别于传统的基于整体特征的融合方法,将局部纹理、颜色、形状等特征作为基础要素,更有利于编码的准确性,从而提升识别性能。 2. 提出一种判别性的视觉字典生成和编码方法,将其与传统的稀疏编码(sparse coding)、局部软编码(local soft coding)和显著编码(salient coding)结合起来,充分利用训练集中的类别信息,在不增加视觉字典规模的情况下,提高其表达能力,进而增加昆虫图像的识别精度。 3. 提出一种基于混合策略的汇聚(pooling)方法,不同于传统的加法汇聚(sum-pooling)对响应的“一视同仁”,也不同于传统的最大汇聚(maxpooling)对响应的“非此即彼”,而是通过对非最大响应的处理,将其所包含的信息汇聚到最后的特征当中,既解决了加法汇聚对于最大响应缺乏优先权的问题,又克服了最大汇聚对非最大响应一概忽略的缺陷,从而更好地利用局部特征来表征整幅昆虫图像。4. 提出一种针对非均匀类别集的分类器融合方法,在对昆虫图像进行分类的实际系统中,昆虫图像常以不同部位照片的形式得到,而对于同一个标本,由于捕捉或保存不当,会存在某个部位缺失的情况,且不同部位的分类器可能与类别集不一致。针对这种情况,我们利用分配-权重矩阵,设计了一种多分类器融合方法,充分利用不同部位图像之间的互补性,提高识别精度。
英文摘要Insect recognition technology has many potential applications such as animal and plant quarantine, agriculture and forestry pests and disease prevention, and ecological system research and protection. However, there is a huge gap between the great demand and the lack of professional entomologists, which becomes a bottleneck of agriculture, forestry, ecology, environmental protection of our country. With the rapid development of computer science and technology, there are significant breakthroughs in computational capabilities, digital image processing technology, computer vision, machine learning, and pattern recognition.Identifying insects using the technology of digital image processing and pattern recognition can compensate for the shortage of entomologists and create enormous economic value for agriculture, forestry, ecology, environmental protection and other industries of our country. To develop automatic insect recognition technology based on computer vision and pattern recognition has significant value from both the academic and application view. With the aim of automatic insect identification, this thesis does a thorough research on image feature extraction, coding method, pooling method and classification of insect images. The main contributions are summarized as follows, 1. We propose a feature level fusion method for insect image recognition. Unlike traditional methods that mostly fuse holistic features, the proposed method fuses three types of local features (local texture, color and shape) before dictionary learning, coding and classification, This can better utilize the discrimination abilities of local features and promote the recognition performance. 2. We propose a discriminative dictionary generation and coding method, and combine with the traditional sparse coding, local soft coding and salient coding. This can make full use of the information of training set, to improve the representation ability of dictionary without increasing its size to increase the insect image recognition accuracy. 3. We propose a hybrid pooling strategy for feature representation. It solves both the problem of the lack of priority issues in sum-pooling and the problem of the ignorance of sub-salient features in max-pooling. This makes better use of the local features to characterize the whole images of insects. 4. We propose a decision fusion method for classifiers of heterogeneous class sets. Practically, we can get images of different parts of insects. For...
语种中文
其他标识符200718014628055
内容类型学位论文
源URL[http://ir.ia.ac.cn/handle/173211/6490]  
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
陆安. 基于视觉的昆虫物种识别算法研究[D]. 中国科学院自动化研究所. 中国科学院大学. 2012.
个性服务
查看访问统计
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。


©版权所有 ©2017 CSpace - Powered by CSpace