Random-valued impulse noise model and L0 loss were added

Author: yusuke-a-uchida

README.md

@@ -110,6 +110,13 @@ If UNet is used, the result becomes 29.67 (noisy targets) vs. 30.14 (clean targe
 
 <img src="result/val_PSNR_text.png" width="480px">
 
+#### Random-valued impulse noise
+<img src="result/val_loss_impulse.png" width="480px">
+
+
+<img src="result/val_PSNR_impulse.png" width="480px">
+
+
 #### Check denoising result
 
 ```bash
@@ -134,6 +141,15 @@ Denoising result by noise target model:
 
 <img src="result/baby_GT_text_noise.png" width="800px">
 
+##### Random-valued impulse noise
+Denoising result by clean target model
+
+<img src="result/baby_GT_impulse_clean.png" width="800px">
+
+Denoising result by noise target model:
+
+<img src="result/baby_GT_impulse_noise.png" width="800px">
+
 
 #### Trained weights
 
@@ -145,6 +161,10 @@ Denoising result by noise target model:
 
 [Text insertion, clean target](https://github.com/yu4u/noise2noise/releases/download/v0.1/weights.056-4.172-28.07752_text_clean.hdf5)
 
+[Random-valued impulse noise, noise target](https://github.com/yu4u/noise2noise/releases/download/v0.1/weights.047-2.317-24.30238_impulse_noise.hdf5)
+
+[Random-valued impulse noise, clean target](https://github.com/yu4u/noise2noise/releases/download/v0.1/weights.038-4.547-24.81654_impulse_clean.hdf5)
+
 ### TODOs
 
 - [x] Compare (noise, clean) training and (noise, noise) training

result/baby_GT_impulse_clean.png

@@ -41,7 +41,7 @@ def get_args():
     parser.add_argument("--steps", type=int, default=1000,
                         help="steps per epoch")
     parser.add_argument("--loss", type=str, default="mse",
-                        help="loss; mse' or 'mae' is expected")
+                        help="loss; mse', 'mae', or 'l0' is expected")
     parser.add_argument("--output_path", type=str, default="checkpoints",
                         help="checkpoint dir")
     parser.add_argument("--source_noise_model", type=str, default="gaussian,0,50",