hananona

Technology

Deep Learning Model

The task here is multi-class classification. Especially it is an example of fine-grained recognition. The number of classes is 792. This number comes from the availability of photos (see Dataset section below). The model used is DenseNet-201, pre-trained by ImageNet dataset.
http://openaccess.thecvf.com/content_cvpr_2017/papers/Huang_Densely_Connected_Convolutional_CVPR_2017_paper.pdf

Dataset

Images of flowers are collected from ImageNet (http://www.image-net.org ). We also added photos of flowers common in Japan. For each of 792 classes, we collected at least 150 images. However, more than half classes have 700 photos. This is a sort of imbalanced dataset, however, by using pretrained model even the classes with less photos could learned well.
Attention: Some images stored in ImageNet are incorrectly labeled. So if you want to use them for your own deep learning experiment, first you need data-cleaning by removing incorrectly labeled images. You may also need to add images to supplement the deleted images. This is what we did. Indeed this is the hardest part of this experiment.

Performance

Using the test dataset that has 50 images for each nodes, the model achieve the following.
top1 accuracy = 86.4%
top5 accuracy = 98.3%