def my_loss3(y_true, y_pred):
MSE_loss = K.mean(K.square(y_pred[:, :, :, :3] - y_true))
SSIM_loss = utls.tf_ssim(
tf.expand_dims(y_pred[:, :, :, 0], -1),
tf.expand_dims(y_true[:, :, :, 0], -1)) + utls.tf_ssim(
tf.expand_dims(y_pred[:, :, :, 1], -1),
tf.expand_dims(y_true[:, :, :, 1], -1)) + utls.tf_ssim(
tf.expand_dims(y_pred[:, :, :, 2], -1),
tf.expand_dims(y_true[:, :, :, 2], -1))
#distortion loss calculated using iou
#resnet_loss = K.mean(K.abs(y_pred[:, :, :, 3:19] - y_pred[:, :, :, 19:35]))
edge_map1 = tf.image.sobel_edges(y_pred[:, :, :, :3])
edge_map2 = tf.image.sobel_edges(y_true)
ssim1 = K.mean(
tf.image.ssim(edge_map1,
edge_map2,
max_val=255,
filter_size=3,
filter_sigma=1.5,
k1=0.01,
k2=0.03))
dist_loss = 1 - ssim1 + K.mean(
K.square(edge_map1 - edge_map2)) + 1 - tf.clip_by_norm(
tf.image.psnr(y_pred[:, :, :, :3], y_true, max_val=1.0), 1)
loss = 3 - SSIM_loss + MSE_loss + dist_loss
return loss
def my_loss(y_true, y_pred):
MAE_loss = K.mean(K.abs(y_pred[:, :, :, :3] - y_true))
SSIM_loss = utls.tf_ssim(
tf.expand_dims(y_pred[:, :, :, 0], -1),
tf.expand_dims(y_true[:, :, :, 0], -1)) + utls.tf_ssim(
tf.expand_dims(y_pred[:, :, :, 1], -1),
tf.expand_dims(y_true[:, :, :, 1], -1)) + utls.tf_ssim(
tf.expand_dims(y_pred[:, :, :, 2], -1),
tf.expand_dims(y_true[:, :, :, 2], -1))
VGG_loss = K.mean(K.abs(y_pred[:, :, :, 3:19] - y_pred[:, :, :, 19:35]))
percent = 0.4
index = int(256 * 256 * percent - 1)
gray1 = 0.39 * y_pred[:, :, :,
0] + 0.5 * y_pred[:, :, :,
1] + 0.11 * y_pred[:, :, :, 2]
gray = tf.reshape(gray1, [-1, 256 * 256])
gray_sort = tf.nn.top_k(-gray, 256 * 256)[0]
yu = gray_sort[:, index]
yu = tf.expand_dims(tf.expand_dims(yu, -1), -1)
mask = tf.to_float(gray1 <= yu)
mask1 = tf.expand_dims(mask, -1)
mask = tf.concat([mask1, mask1, mask1], -1)
low_fake_clean = tf.multiply(mask, y_pred[:, :, :, :3])
high_fake_clean = tf.multiply(1 - mask, y_pred[:, :, :, :3])
low_clean = tf.multiply(mask, y_true[:, :, :, :])
high_clean = tf.multiply(1 - mask, y_true[:, :, :, :])
Region_loss = K.mean(
K.abs(low_fake_clean - low_clean) * 4 +
K.abs(high_fake_clean - high_clean))
loss = MAE_loss + VGG_loss / 3. + 3 - SSIM_loss + Region_loss
return loss
def my_loss(y_true, y_pred):
MSE_loss = K.mean(K.abs(y_pred[:, :, :, :3] - y_true))
SSIM_loss = utls.tf_ssim(
tf.expand_dims(y_pred[:, :, :, 0], -1),
tf.expand_dims(y_true[:, :, :, 0], -1)) + utls.tf_ssim(
tf.expand_dims(y_pred[:, :, :, 1], -1),
tf.expand_dims(y_true[:, :, :, 1], -1)) + utls.tf_ssim(
tf.expand_dims(y_pred[:, :, :, 2], -1),
tf.expand_dims(y_true[:, :, :, 2], -1))
#psnr=tf.clip_by_norm(tf.image.psnr(y_pred[:,:,:,:3], y_true,max_val=1.0),1)
return 3 - SSIM_loss + 3 * MSE_loss