›› 2018, Vol. 61 ›› Issue (2): 255-262.doi: 10.16380/j.kcxb.2018.02.013

• 研究论文 • 上一篇    下一篇

基于全卷积网络的鳞翅目标本图像前背景自动分割方法研究

竺乐庆1,*, 马梦园1, 张真2, 孟昭军3, 吴伟4, 任利利5, 高翠青6, 南小宁7   

  1. (1.浙江工商大学计算机与信息工程学院, 杭州 310018; 2. 中国林业科学研究院森林生态与保护研究所, 国家林业局森林保护学重点实验室, 北京 100091; 3. 东北林业大学林学院, 哈尔滨 150040; 4. 西南林业大学保护生物学学院, 昆明 650224; 5. 北京林业大学林学院, 北京100083; 6. 南京林业大学林学院, 南京 210037; 7. 西北农林科技大学林学院, 陕西杨凌 712100)
  • 出版日期:2018-02-20 发布日期:2018-02-20

Foreground-background segmentation of lepidopteran specimen images based on fully convolutional networks

ZHU Le-Qing1,*, MA Meng-Yuan1, ZHANG Zhen2, MENG Zhao-Jun3, WU Wei4, REN Li-Li5, GAO Cui-Qing6, NAN Xiao-Ning7   

  1. (1. School of Computer Science and Information Engineering, Zhejiang Gongshang University, Hangzhou 310018, China; 2. Key Laboratory of Forest Protection of State Forestry Administration, Research Institute of Forest Ecology, Environment and Protection, Chinese Academy of Forestry Sciences, Beijing 100091, China; 3. School of Forestry, Northeast Forestry University, Harbin 150040, China; 4. Faculty of Conservation Biology, Southwest Forestry University, Kunming 650224, China; 5. School of Forestry, Beijing Forestry University, Beijing 100083, China; 6. School of Forestry, Nanjing Forestry University, Nanjing 210037, China; 7. College of Forestry, Northwest A&F University, Yangling, Shaanxi 712100, China)
  • Online:2018-02-20 Published:2018-02-20

摘要: 【目的】本研究旨在探索使用计算机视觉技术实现对鳞翅目标本图像的前背景分割方法。【方法】首先对用于训练和测试的昆虫标本图像去除背景,获得昆虫图像的前背景分割参考标准,对过大的昆虫图像进行缩小处理;其次对训练集图像采用旋转、平移、缩放等方法进行数据增强,剪切出中心区域作为有效图像。求取所有训练样本的均值图像,并从所有输入中减去该均值图像。测试用图像只做归一化但不进行数据增强。微调全卷积神经网络,重点调整结构产生变化的卷积层和反卷积层的参数,用前述训练数据集训练直至收敛。对于待分割图像,只要将图像归一化后输入到训练好的全卷积网络,网络将输出前背景分割结果。【结果】该方法在包含823个样本的测试集中进行了测试,取得的mIoU(meanIntersection over Union)达94.96%,而且分割的视觉效果已经非常接近于人工分割的结果。【结论】实验结果证明通过训练全卷积神经网络可以有效实现鳞翅目标本图像的前背景自动分割。

关键词: 鳞翅目, 图像处理, 前背景分割, 深度学习, 全卷积神经网络

Abstract:  【Aim】 This study aims to realize the automatic foreground-background segmentation of lepidopteran specimen images by exploring the state-of-art computer vision technology. 【Methods】 First, the background is manually removed to form the ground truth of training set and testing set, and those images that are too large are resized to smaller ones. Then, the training set is enhanced by rotation, translation, scaling, etc., and their central areas are cropped as valid input and target images. Afterwards, the mean image of all the training samples is calculated and subtracted from all input images. Testing images are simply normalized but not enhanced. Fully convolutional networks (FCNs) are fine-tuned with training set until they converge. The parameter adjustment on later convolutional layers and de-convolutional layers is emphasized since their structures are different from those of original immigrated FCNs. When one given insect image is fed into the trained FCN after normalization, the segmentation result will be given. 【Results】 The proposed method was evaluated with the testing set including 823 samples, and the final mIoU (mean Intersection over Union) was as high s 94.96%. The visual effect of segmentation results given by FCN was much close to the manually produced results.【Conclusion】 The experimental results prove that the foregroundbackground of lepidopteran specimen images can be segmented efficiently by the trained FCN.

Key words: Lepidoptera, image processing, foreground-background segmentation, deep learning, fully convolutional network (FCN)