Image Retrieval based on Binary Hash Code and CNN

Image Retrieval realizing cvprw15 paper，including Face image and other image

Created by zhleternity

This repo is based on the cvprw15 paper about image retrieval(proposed by Kevin Lin - cvprw15) for CBIR tasks.it is about the using of binary hash code with image retrieval.

• The extraction of binary hash codes & deep features
• Fast indexing of both binary hash codes & deepfeatures
• Fast computing of similarity (distances) based on features & binary codes extrated
• Easy request for similar images (in database)
• Sorted visualization of results

In the repo, you can test on Webface dataset and foods25 dataset directly. If you wanna to download the above mentioned datasets and model,please have a look on Running this demo section to downlowd what you wanted. But also,you can train your own dataset and other datasets from scratch,and then do the testing.Please have a look on Training From Scratch section to do the training work.

OS: Ubuntu 16.04 64 bit
GPU: Nvidia RTX 2070
Cuda 9.2
CuDNN 7.5.0
Python 2.7.12
OpenCV 3.2.0

1. Caffe(https://github.com/BVLC/caffe)
2. Python's packages (requirements.txt),install command is: $ pip install -r requirements.txt
3. OpenCV((http://www.pyimagesearch.com/2016/10/24/ubuntu-16-04-how-to-install-opencv/)

1. Download this repository:

   $ git clone --recursive https://github.com/zhleternity/image_retrieval_binary_hash_code.git
   $ cd image-retrieval-lsh
   

   ├── corefuncs #main functions about fetaure extraction,indexing,distances,and results showing
   ├── data  #dataset and extracted feature file(.hdf5)
   ├── examples #model and prototxt
   ├── results #output results
   |── indexing.py #extract features,and build the index
   ├── retrieve.py #do the image retrieval
   ├── init_path.py #init caffe path and other config
   ├── README.md
   └── requirements.txt

2. Download datasets and model:

For the foods25 dataset,it has already exists in this repo,so you need not downlowd. For the Webface dataset, you need download by yourself. Get the data and models required:

• Download the data and uncompress in 'image_retrieval_binary_hash_code/data'

  • foods25: Google Drive

  • Webface: http://www.cbsr.ia.ac.cn/english/CASIA-WebFace-Database.html

• Download the models：

  For foods25,uncompress in 'image_retrieval_binary_hash_code/examples/'

  • foods25: Google Drive

  For webface,download the model and put it in 'image_retrieval_binary_hash_code/examples/facescrub/models/'

  • Webface: https://pan.baidu.com/s/1ecbDJgp48IvRBrU4RHTFKA Extract Code：p5l3

3. Extract features && Build the index:

$ python indexing.py -u <use GPU or not,0 or 1> -d <path to your dataset> -o <your output data-db name>

4. Searching:

$ python retrieve.py -u <use GPU or not,0 or 1> -f <your output data-db name> -d <path to your dataset> -q <your query image>

1. Feature extraction & indexing of a sample dataset (foods25) with the following commands:

   $ python indexing -u 1 -d data/foods25/imgs -o data/foods25/foods25.hdf5
   

   In my repo, the foods25.hdf5 have already exists,so you can skip thie step.

The output of this command is stored as 'data/foods25/foods25.hdf5'.

2. Search for similar images:

   $ python retrieve.py -u 1 -f data/foods25/foods25.hdf5 -d data/foods25/imgs -q data/foods25/imgs/taro/28581814.jpg
   

The output of this command is stored in 'results/'

When you run the above Experiments section,the model and prototxt path in indexing.py && retrieve.py should be corresponding to your dataset.

Please see th following screenshot:

Query:

Results:

Now,we begin to train the demo from scratch.And then you can test the retrieval on your trained model.

• Split your data

$ cd data/facescrub

$ mkdir train && mkdir val

$ python split_dataset.py

Before this op,you will see your images is in train && val

• Convert txt

$ touch train.txt && touch val.txt

$ python generate_txt.py

This operation will yeild two non-empty txt profile:train.txt && val.txt

• Convert data into LMDB or LevelDB

$ sh create_imagenet.sh

This operation will yeild two leveldb profile:facescrub_train_leveldb && facescrub_val_leveldb

Attention that:you may need do some change on the path,please change it to correspond to your directory path and dataset.

• Train

$ cd ../../examples/facescrub

Please replace the data source path to your leveldb path in train_face_48.prototxt and test_face_48.prototxt .Also change the path in solver_face_48.prototxt.

$ /xxx/xxx/caffe/build/tools/caffe train -solver solver_face_48.prototxt -gpu 1

All this steps are easy for human who are familiar to Caffe，hence issues above mentioned are brief.

Please pull issues to me directly,i will return you on time.