使用高效超分辨率的图像恢复。凯拉斯，张量流_Python_下载.zip

# SR-Restore Super resolution and Image restoration. ## Install Requires `tensorflow==1.1.0` and `keras==2.0.2` $ pip install -r ./requirements.txt ## Super Resolution example ```` # Resize low-res and high-res images for Super Resolution $ python -m sr.images data/images/raw/ \ --lr data/images/lr \ --hr data/images/hr \ --lr_size 72x72 \ --hr_size 288x288 # Train Super Resolution model. # Expects images in /path/images/lr/ and /path/images/hr/ $ python -m sr.train data/images \ --results results/ \ --network espcnn \ --scale 4 \ --epochs 100 \ --batch_size 32 # Predict high-res images using trained model weights. # (Optional) use `--limit` option to limit num of images to perform on. $ python -m sr.predict data/images/lr/ \ --output results/output/ \ --weights results/espcnn_20170701_190534/weights.h5 \ --network espcnn \ --scale 4 \ --batch_size 32 \ --limit 20 ```` ## Image Restore example ```` # Resize a directory of images to single size # We need /image_path/original/ and /image_path/enhanced/ $ python -m restore.images data/images/raw_original \ --output data/images/original \ --size 288x288 $ python -m restore.images data/images/raw_enhanced \ --output data/images/enhanced \ --size 288x288 # Train Restore model. # Expects images in /path/images/original/ and /path/images/enhanced/ # Saves model weights, checkpoints, and tensorboard logs to path specified in results. $ python -m sr.train data/images \ --results results/ \ --network cnn \ --epochs 100 \ --batch_size 32 # Predict original images using trained model weights. # (Optional) use `--limit` option to limit num of images to perform on. $ python -m sr.predict data/images/enhanced/ \ --output results/output/ \ --weights results/cnn_20170701_190534/weights.h5 \ --network cnn \ --batch_size 32 \ --limit 20 ```` ## Dataset celebA: (link) ## References - Image Super Resolution (link) - Perceptual loss (link) - Efficient Real time convolution (link)