英语原文共 5 页，剩余内容已隐藏，支付完成后下载完整资料
Large-Scale Plant Classification with Deep Neural Networks
Ignacio Heredialowast;
Instituto de Fisica de Cantabria (CSIC-UC)
E-39005, Santander, Spain
ABSTRACT
This paper discusses the potential of applying deep learning techniques for plant classification and its usage for citizen science in large-scale biodiversity monitoring. We show that plant classification using near state-of-the-art convolutional network architectures like ResNet50 achieves significant improvements in accuracy compared to the most widespread plant classification application in test sets composed of thousands of different species labels. We find that the predictions can be confidently used as a baseline classification in citizen science communities like iNaturalist (or its Spanish fork, Natusfera) which in turn can share their data with biodiversity portals like GBIF.
KEYWORDS
deep learning, plant classification, citizen science, biodiversity monitoring
ACM Reference format:
Ignacio Heredia. 2017. Large-Scale Plant Classification with Deep Neural Networks. In Proceedings of ACM CFrsquo;17, Siena, Italy, May 15-17, 2017, 5 pages. DOI: 10.1145/3075564.3075590
- INTRODUCTION
The deep learning revolution has brought significant advances in a number of fields [1], primarily linked to image and speech recognition. The standardization of image classification tasks like the ImageNet Large Scale Visual Recognition Challenge [2] has resulted in a reliable way to compare top performing architectures. Since the AlexNet architecture [3], the first efficient implementation of convolutional neural networks using GPUs, the error in these competitions has reached superhuman performance [4].
Despite this recent success in general image recognition, the work in the biodiversity community relies heavily on hand labeled image data assigned by a (relatively) small community of experts and does not exploit these recent advances. This might be an impediment to open the community to non expert users who, armed with modern technologies handily embedded in a smartphone, can push biodiversity monitoring to the next level. The use of deep learning for plant classification is not novel [5, 6] but has mainly focused in leaves and has been restricted to a limited amount of species, therefore making it of limited use for large-scale biodiversity monitoring purposes. This same specificity issue applies to some standardized plant datasets [7] which are very helpful to
lowast;Electronic address: iheredia@ifca.unican.es
ACM CFrsquo;17, Siena, Italy
copy; 2017 Copyright held by the owner/author(s). Publication rights licensed to ACM. This is the authorrsquo;s version of the work. It is posted here for your personal use. Not for redistribution. The definitive Version of Record was published in Proceedings of ACM CFrsquo;17, May 15-17, 2017, https://doi.org/10.1145/3075564.3075590.
evaluate the network performances but who are limited in variety of species or in the diversity of the images (focusing mainly in flowers or leaves). The PlantNet tool [8, 9], based on distant versions of the IKONA algorithms, pioneered in creating an open access tool to automate the task of recognizing a wide variety of species. However it does not reach the performance of expert botanists. Applying the recent advances in convolutional neural networks could have a positive impact in closing this performance gap. This could be a large step towards building a reliable and general large-scale plant recognition app that spreads the use of citizen science for biodiversity monitoring.
- THE DATASET
As training dataset we use the great collection of images which are available in PlantNet under a Creative-Common AttributionShareAlike 2.0 license. It consists of around 250K images belonging to more than 6K plant species of Western Europe. These species are distributed in 1500 genera and 200 families. Each image has been labeled by experts and comes with a tag which specifies the focus of the image, like rsquo;habitrsquo;, rsquo;flowerrsquo;, rsquo;leafrsquo;, rsquo;barkrsquo;, etc. Most images have resolutions ranging from 200K to 600K pixels and aspect ratios ranging from 0.5 to 2. The dataset is highly unbalanced because most labels contain very few images.
We train on the whole dataset (without making validation or test splits) as we intend to build a classifier trained on the same dataset as the PlantNet tool so that their performances can be fairly compared. Also we believe that testing the classification performance on a subset of PlantNet is not an accurate measure of the performance of the net on real-world data as all the images in the dataset are highly correlated (many photos inside a specie share author and are often taken from the same plant with slightly different angles). Therefore at test time we will use three external datasets to confidently measure the performance of our net.
- THE MODEL
As plant classification is not very different from general object classification, we expect that top performing architectures in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) would perform well in this task. Therefore we use as convolutional neural network architecture the ResNet model [10] who won the ILSVRCrsquo;15. This architecture consists of a stack of similar (so-called residual) blocks, each block being in turn a stack of convolutional layers. The innovation is th
剩余内容已隐藏，支付完成后下载完整资料
资料编号：[254009]，资料为PDF文档或Word文档，PDF文档可免费转换为Word
以上是毕业论文外文翻译，课题毕业论文、任务书、文献综述、开题报告、程序设计、图纸设计等资料可联系客服协助查找。
