diff --git a/dan/ocr/transforms.py b/dan/ocr/transforms.py
index 8a10ce170857a284b43715972e4a0e1c3580e1b6..3a8c74f64ac5f586a02ce070f4ca6aaf1c3bd03f 100644
--- a/dan/ocr/transforms.py
+++ b/dan/ocr/transforms.py
@@ -6,6 +6,7 @@ from enum import Enum
from random import randint
import albumentations as A
+import cv2
import numpy as np
from albumentations.augmentations import (
Affine,
@@ -15,16 +16,16 @@ from albumentations.augmentations import (
GaussianBlur,
GaussNoise,
Perspective,
+ RandomScale,
Sharpen,
ToGray,
)
from albumentations.core.transforms_interface import ImageOnlyTransform
-from cv2 import dilate, erode, resize
+from cv2 import dilate, erode
from numpy import random
from torch import Tensor
-from torch.distributions.uniform import Uniform
from torchvision.transforms import Compose, ToPILImage
-from torchvision.transforms.functional import resize as resize_tensor
+from torchvision.transforms.functional import resize
class Preprocessing(str, Enum):
@@ -54,7 +55,7 @@ class FixedHeightResize:
def __call__(self, img: Tensor) -> Tensor:
size = (self.height, self._calc_new_width(img))
- return resize_tensor(img, size, antialias=False)
+ return resize(img, size, antialias=False)
def _calc_new_width(self, img: Tensor) -> int:
aspect_ratio = img.shape[2] / img.shape[1]
@@ -71,7 +72,7 @@ class FixedWidthResize:
def __call__(self, img: Tensor) -> Tensor:
size = (self._calc_new_height(img), self.width)
- return resize_tensor(img, size, antialias=False)
+ return resize(img, size, antialias=False)
def _calc_new_height(self, img: Tensor) -> int:
aspect_ratio = img.shape[1] / img.shape[2]
@@ -96,7 +97,7 @@ class MaxResize:
ratio = min(height_ratio, width_ratio)
new_width = int(width * ratio)
new_height = int(height * ratio)
- return resize_tensor(img, (new_height, new_width), antialias=False)
+ return resize(img, (new_height, new_width), antialias=False)
class Dilation:
@@ -156,29 +157,6 @@ class ErosionDilation(ImageOnlyTransform):
)
-class DPIAdjusting(ImageOnlyTransform):
- """
- Resolution modification
- """
-
- def __init__(
- self,
- min_factor: float = 0.75,
- max_factor: float = 1,
- always_apply: bool = False,
- p: float = 1.0,
- ):
- super(DPIAdjusting, self).__init__(always_apply, p)
- self.min_factor = min_factor
- self.max_factor = max_factor
- self.p = p
- self.always_apply = False
-
- def apply(self, img: np.ndarray, **params):
- factor = float(Uniform(self.min_factor, self.max_factor).sample())
- return resize(img, None, fx=factor, fy=factor)
-
-
def get_preprocessing_transforms(
preprocessings: list, to_pil_image: bool = False
) -> Compose:
@@ -212,7 +190,10 @@ def get_augmentation_transforms() -> A.Compose:
"""
return A.Compose(
[
- DPIAdjusting(min_factor=0.75, max_factor=1),
+ # Scale between 0.75 and 1.0
+ RandomScale(
+ scale_limit=[-0.25, 0], always_apply=True, interpolation=cv2.INTER_AREA
+ ),
A.SomeOf(
[
ErosionDilation(min_kernel=1, max_kernel=4, iterations=1),
@@ -220,10 +201,18 @@ def get_augmentation_transforms() -> A.Compose:
GaussianBlur(sigma_limit=2.5, p=1),
GaussNoise(var_limit=50**2, p=1),
ColorJitter(
- contrast=0.2, brightness=0.2, saturation=0.2, hue=0.2, p=1
+ contrast=0.2,
+ brightness=0.2,
+ saturation=0.2,
+ hue=0.2,
+ p=1,
),
ElasticTransform(
- alpha=20.0, sigma=5.0, alpha_affine=1.0, border_mode=0, p=1
+ alpha=20.0,
+ sigma=5.0,
+ alpha_affine=1.0,
+ border_mode=0,
+ p=1,
),
Sharpen(alpha=(0.0, 1.0), p=1),
Affine(shear={"x": (-20, 20), "y": (0, 0)}, p=1),
diff --git a/docs/assets/augmentations/document_original.png b/docs/assets/augmentations/document_original.png
new file mode 100644
index 0000000000000000000000000000000000000000..57fe507d33f2ee10c831c6c5bf7fa789904d110b
Binary files /dev/null and b/docs/assets/augmentations/document_original.png differ
diff --git a/docs/assets/augmentations/document_random_scale.png b/docs/assets/augmentations/document_random_scale.png
new file mode 100644
index 0000000000000000000000000000000000000000..1da83f90866f7fa1393f1f0f6419b987eabe9e68
Binary files /dev/null and b/docs/assets/augmentations/document_random_scale.png differ
diff --git a/docs/assets/augmentations/line_full_pipeline_2.png b/docs/assets/augmentations/line_full_pipeline_2.png
new file mode 100644
index 0000000000000000000000000000000000000000..ca0e9203977dd46c28e86548b69bf6f75dab96fd
Binary files /dev/null and b/docs/assets/augmentations/line_full_pipeline_2.png differ
diff --git a/docs/assets/augmentations/line_original.png b/docs/assets/augmentations/line_original.png
new file mode 100644
index 0000000000000000000000000000000000000000..46d9cc3421ad47fda6fc12791296c037b4096c74
Binary files /dev/null and b/docs/assets/augmentations/line_original.png differ
diff --git a/docs/assets/augmentations/line_random_scale.png b/docs/assets/augmentations/line_random_scale.png
new file mode 100644
index 0000000000000000000000000000000000000000..1c402e682ec6c7fe7b687be7ceac837bf0d4f4bd
Binary files /dev/null and b/docs/assets/augmentations/line_random_scale.png differ
diff --git a/docs/usage/train/augmentation.md b/docs/usage/train/augmentation.md
index 1f98e0244ca9346b0f27ecdeac21c8ee4baff277..cc18d45a797362dc356d56b74aa47ebff87cb8ea 100644
--- a/docs/usage/train/augmentation.md
+++ b/docs/usage/train/augmentation.md
@@ -107,14 +107,16 @@ This page lists data augmentation transforms used in DAN.
| Examples |   |
| CPU time (seconds/10 images) | 0.02 (3013x128 pixels) / 0.02 (1116x581 pixels) |
-### DPIAdjusting
+### Random Scale
-| | DPIAdjusting |
-| ----------- | -------------------------------------------------------------- |
-| Description | This transformation downscales the image from a random factor. |
-| Comments | Similar to the original DAN implementation. |
+| | RandomScale |
+| ------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| Description | This transformation downscales the image from a random factor. |
+| Comments | The original DAN implementation reimplemented it as [DPIAdjusting](https://github.com/FactoDeepLearning/DAN/blob/da3046a1cc83e9be3e54dd31a5e74d6134d1ebdc/basic/transforms.py#L62). |
+| Documentation | See the [`albumentations` documentation](https://albumentations.ai/docs/api_reference/augmentations/geometric/resize/#albumentations.augmentations.geometric.resize.RandomScale) |
+| Examples |   |
-### ToGray
+### To Gray
| | ToGray |
| ---------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------- |
diff --git a/docs/usage/train/config.md b/docs/usage/train/config.md
index 148f0fa96536b453afd2337607d46063c4ef402e..a435fdc6340f2e8a9ed3987021c0dfc00eb24724 100644
--- a/docs/usage/train/config.md
+++ b/docs/usage/train/config.md
@@ -134,7 +134,8 @@ The following configuration is used by default when using the `teklia-dan train`
```py
transforms = A.Compose(
[
- DPIAdjusting(min_factor=0.75, max_factor=1),
+ # Scale between 0.75 and 1.0
+ RandomScale(scale_limit=[-0.25, 0], always_apply=True, interpolation=cv2.INTER_AREA),
A.SomeOf(
[
ErosionDilation(min_kernel=1, max_kernel=4, iterations=1),
diff --git a/tests/data/training/models/best_0.pt b/tests/data/training/models/best_0.pt
index 79bcb28a2388f266d497ee4d612f1835eba5cae6..889355b039f64ff270086a11714986e13621a50b 100644
--- a/tests/data/training/models/best_0.pt
+++ b/tests/data/training/models/best_0.pt
@@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
-oid sha256:428ceb4d08363c05b6e60e87e5e1ae65560d345756926c23f13e6d191dc33d69
+oid sha256:a1fc3f9bdf52055aadefe49ff3537d9e6c9b1af5b14364d027d3fcdebeabf9a7
size 84773087
diff --git a/tests/data/training/models/last_3.pt b/tests/data/training/models/last_3.pt
index c7ecb5dbaa3f26c8d0933faa42685277b66cbd05..2cbaf1a2cf52e11e10390e51c98969d01de67038 100644
--- a/tests/data/training/models/last_3.pt
+++ b/tests/data/training/models/last_3.pt
@@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
-oid sha256:c2029a5822c5a8d4253a95c33a05357e707b9b46d056988dcab730945dfd5775
+oid sha256:8e8536dc913af4d1560413c1d91aa41d1636c0b8ed557389b82e74fb991312b0
size 84773087