New Zealand pollen

Dataset first used in the paper <i>Precise automatic classification of 46 different pollen types with convolutional neural networks</i> [0].<br><br>The images are dark field microscope images captured on the Classifynder system. The Classifynder (formerly knownas AutoStage) is documented in [1]. It was designed as a complete ’standalone’ system for automated pollen analysis. The system uses basic shape features to identify the locations of objects of interest (i.e. pollen grains) in conventional microscope slides under a low-power objective. It then switches to a higher power objective and visits the location of each pollen object to capture an image of it to be used for classification.<br>Objects are imaged at different focus levels, producing a ’Z-stack’. The system subsets the best-focused portions of each object Z-stack images, and then combines them to<br>produce a single composite image, followed by segmentation from the background. This image is then falsely colored to show depth.<br>[0] http://doi.org/10.1371/journal.pone.0229751[1] http://www.sciencedirect.com/science/article/pii/S0034666711001205

本数据集首次在论文《基于卷积神经网络的46种花粉类型精准自动分类》（Precise automatic classification of 46 different pollen types with convolutional neural networks）[0]中使用。

该数据集的图像为在Classifynder系统（原名为AutoStage）上采集的暗视野显微镜图像。该系统的相关技术文档可参见文献[1]，其被设计为一套完整的独立式自动化花粉分析系统。系统首先通过基础形状特征，识别低倍物镜下常规显微镜载玻片上的目标物体（即花粉粒）的位置；随后切换至高倍物镜，逐一定位各花粉目标并采集图像，用于后续分类任务。

系统会在不同对焦层级下对目标物体成像，生成Z堆叠（Z-stack）图像。随后，系统会截取每个物体Z堆叠图像中对焦最佳的区域，将这些区域拼接为单张合成图像，再完成背景分割。最后，对该图像进行伪彩色着色处理以展示景深信息。

[0] http://doi.org/10.1371/journal.pone.0229751
[1] http://www.sciencedirect.com/science/article/pii/S0034666711001205