A modified version of morimoris's SRCNN implement based on the paper called Image Super-Resolution Using Deep Convolutional Networks.
Now the program can output colorful images, but that don't mean this program can really process the color part of input images. In fact, I just keep the Cr and Cb color spaces and add them to the output images which only have Y color space to make them colorful. So, there is no essential difference between this program and the original one about the process of color spaces.
The processing effect is roughly as shown in the figure:
 |
 |
|---|---|
| Low Resolution | SRCNN |
About the evaluation, SSIM is included now.
- cuda
- tensorflow
- numpy
- opencv-python
The detail info about all the args are shown below. You can use the default value or use specified value. Please use --mode to specify a certain mode instead of using default train_model mode.
Generally speaking, the order in which the patterns are used is train_data_create, test_data_create, train_model, evaluate.
The code logic is pretty simple. If you have any confusion, you could check the code for details, or check the original author’s blog. Also an issue is welcomed :)
usage: main.py [-h] [--train_height TRAIN_HEIGHT] [--train_width TRAIN_WIDTH] [--test_height TEST_HEIGHT]
[--test_width TEST_WIDTH] [--train_dataset_num TRAIN_DATASET_NUM] [--test_dataset_num TEST_DATASET_NUM]
[--train_cut_num TRAIN_CUT_NUM] [--test_cut_num TEST_CUT_NUM] [--train_path TRAIN_PATH] [--test_path TEST_PATH]
[--learning_rate LEARNING_RATE] [--BATCH_SIZE BATCH_SIZE] [--EPOCHS EPOCHS] [--mode MODE]
Tensorflow SRCNN Example
optional arguments:
-h, --help show this help message and exit
--train_height TRAIN_HEIGHT
Train data size(height)
--train_width TRAIN_WIDTH
Train data size(width)
--test_height TEST_HEIGHT
Test data size(height)
--test_width TEST_WIDTH
Test data size(width)
--train_dataset_num TRAIN_DATASET_NUM
Number of train datasets to generate
--test_dataset_num TEST_DATASET_NUM
Number of test datasets to generate
--train_cut_num TRAIN_CUT_NUM
Number of train data to be generated from a single image
--test_cut_num TEST_CUT_NUM
Number of test data to be generated from a single image
--train_path TRAIN_PATH
The path containing the train image
--test_path TEST_PATH
The path containing the test image
--learning_rate LEARNING_RATE
Learning_rate
--BATCH_SIZE BATCH_SIZE
Training batch size
--EPOCHS EPOCHS Number of epochs to train for
--mode MODE train_data_create, test_data_create, train_model, evaluate
By the way, please don't forget to get the datasets and put them in a proper place before you run the code.
All the changes have been authorized by the original author.
Original author's Blog：DeepLearningを用いた超解像手法/SRCNNの実装
Original Repo：morimoris/keras_SRCNN
license：morimoris/keras_SRCNN#4