The benchmark collection named Similar-Document-Image-Retrieval-Dataset(SDIRD) is provided for a task, which is finding out the similar document images for a query document image. The structure of SDIRD is as follows:
├── database
├── databaseClassified
├── queryset
├── trainingset
└── database-subject
This is the document image database with 10240 document images.
This directory collects the images of above database according to corresponding original source(148 papers).
This is the query image dataset with 1000 images totally.
This is the target dataset to fine-tune pre-trained CNN models, which including training set with 1000 document images and validation set with 200 images, and the label or category information.
This is the information of document image database about the category id, category label, the number of images from corresponding paper and title.
| Methods | Top-1(%) | Top-3(%) | Top-5(%) | Top-10(%) |
|---|---|---|---|---|
| GoogLeNet | 8.3 | 13.8 | 17.3 | 21.6 |
| ResNet-152 | 17.2 | 24.7 | 28.6 | 35.2 |
| AlexNet | 32.6 | 45.0 | 51.6 | 59.6 |
| Fine-tuned-AlexNet | 37.5 | 54.1 | 60.9 | 69.8 |
| VGGNet-D | 22.8 | 34.0 | 39.3 | 46.9 |
| Fine-tuned-VGGNet-D | 37.2 | 56.2 | 64.6 | 76.5 |
| VGGNet-E | 24.1 | 37.5 | 44.1 | 53.2 |
| Fine-tuned-VGGNet-E | 46.8 | 68.9 | 79.3 | 90.0 |
| Ours | 54.1 | 82.9 | 97.3 | 100.0 |
| Ours | 58.8 | 86.4 | 94.5 | 98.9 |