Remote sensing image retrieval

this is an implement of DenseNet using keras ,this project can do Remote sensing image classifiy or retrieval.I trained and evaluated this model on a dataset called PatternNet.

Dependencies

The project was tested in the following environment

• python 3.5.2
• Keras 2.1.6
• h5py
• tensorflow-gpu 1.9.0
• pillow(PIL) 5.0.0
• numpy 1.14.0
• sklearn

I build this project in Windows 10 so there maybe some path problems when rebuild it in linux/MacOSX.

You may also need a GPU to speed up the train process,so installtion of the CUDA/CUDNN kit maybe necessary.

How to use

Data setup

Using my PatternNet dataset or your own dataset,all you need to do is

• Create a new folder,such as

myDataset

• for each catagory,Create a new folder in 'myDataset'

myDataset/

catagory1 catagory2 catagory3 ...

• put the image into these folders

After this,you need to run utils.write_csv().This function need 3 parameters image_folder, csv_train_path, csv_test_path.This function automatically divides the data into training and test set,and the test set is 30 percent of the total dataset.

import utils

image_folder = r'PatternNet'
csv_train_path = r'PatternNet_train.csv'
csv_test_path = r'PatternNet_test.csv'
	
write_csv(image_folder, csv_train_path, csv_test_path)

Models

To use pretrained weights,you should set the weights='imagenet' when reference DenseNet.DenseNet()

When training your own dataset from scratch,simply set the weights=None.

Train

To train a model use

from train_and_val import train

lr = 1e-6
epochs = 70
batch_size = 16
classes = 38
image_size = 224
model_path = 'pattern.h5'
log_filepath = 'pattern_log'
csv_train_path = r'PatternNet_train.csv'
csv_test_path = r'PatternNet_test.csv'

train(image_size,classes,
      csv_train_path,csv_test_path,
      batch_size,epochs,lr,
      log_filepath,model_path)

The parameter above can be changed to adjust model performance.

here is the weights file -pattern.h5

After training,you can call tensorboard to see the results in detail.use cd command to navigate to the project folder,then

>tensorboard --logdir log_filepath

Image classification

To do classification job,simply use the weights file that you trained in the last step,and call train_and_val.validate().Here are the demo:

from train_and_val import validate

image_size = 224
model_path = 'pattern.h5'
classes = 38
csv_val_path = r'PatternNet_test.csv'
csv_result_path = r'PatternNet_result.csv'

validate(image_size,classes,
          model_path,csv_val_path,csv_result_path)

The result is saved csv_result_path.

Image retrieval

This function contains two parts: index and retrieval

index

This is an implement of feature extraction.The DenseNet's avg pool layer is used to extract features .img_path,label,features are saved into index_file.

from train_and_val import index

image_size = 224
model_path = 'pattern.h5'
classes = 38
csv_imageLib_path = r'PatternNet_test.csv'
index_file = r'PatternNet_index.h5'

index(image_size,classes,
      model_path,csv_imageLib_path,index_file)

here is the index-file PatternNet_index.h5

retrieval

This function use Euclidean distance to find out the image which is similar as target image.Note that target_path need to be a csv file which formatted as image_path,category,only one image should be in this file

from train_and_val import retrieval

image_size = 224
model_path = 'pattern.h5'
classes = 38
target_path = r'target.csv'
retrieval_result_file = r'retrieval_result.txt'

retrieval(image_size,classes,
			model_path,target_path,index_file,
			retrieval_result_file)

Other things

This is my second project in Github,hope you can star or fork this project~

Any issue,you can contact me on QQ 2043494361 or just email me at 204349461@qq.com.

Hope you enjoy this~