4 files changed, 170 insertions, 0 deletions
diff --git a/text_recognizer/data/stems/__init__.py b/text_recognizer/data/stems/__init__.py
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/text_recognizer/data/stems/__init__.py
diff --git a/text_recognizer/data/stems/image.py b/text_recognizer/data/stems/image.py
new file mode 100644
index 0000000..f04b3a0
--- /dev/null
+++ b/text_recognizer/data/stems/image.py
@@ -0,0 +1,18 @@
+from PIL import Image
+import torch
+from torch import Tensor
+import torchvision.transforms as T
+
+
+class ImageStem:
+    def __init__(self) -> None:
+        self.pil_transform = T.Compose([])
+        self.pil_to_tensor = T.ToTensor()
+        self.torch_transform = torch.nn.Sequential()
+
+    def __call__(self, img: Image) -> Tensor:
+        img = self.pil_transform(img)
+        img = self.pil_to_tensor(img)
+        with torch.no_grad():
+            img = self.torch_transform(img)
+        return img
diff --git a/text_recognizer/data/stems/line.py b/text_recognizer/data/stems/line.py
new file mode 100644
index 0000000..2fe1a2c
--- /dev/null
+++ b/text_recognizer/data/stems/line.py
@@ -0,0 +1,86 @@
+import random
+
+from PIL import Image
+import torchvision.transforms as T
+
+import text_recognizer.metadata.iam_lines as metadata
+from text_recognizer.data.stems.image import ImageStem
+
+
+class LineStem(ImageStem):
+    """A stem for handling images containing a line of text."""
+
+    def __init__(
+        self, augment=False, color_jitter_kwargs=None, random_affine_kwargs=None
+    ):
+        super().__init__()
+        if color_jitter_kwargs is None:
+            color_jitter_kwargs = {"brightness": (0.5, 1)}
+        if random_affine_kwargs is None:
+            random_affine_kwargs = {
+                "degrees": 3,
+                "translate": (0, 0.05),
+                "scale": (0.4, 1.1),
+                "shear": (-40, 50),
+                "interpolation": T.InterpolationMode.BILINEAR,
+                "fill": 0,
+            }
+
+        if augment:
+            self.pil_transforms = T.Compose(
+                [
+                    T.ColorJitter(**color_jitter_kwargs),
+                    T.RandomAffine(**random_affine_kwargs),
+                ]
+            )
+
+
+class IamLinesStem(ImageStem):
+    """A stem for handling images containing lines of text from the IAMLines dataset."""
+
+    def __init__(
+        self, augment=False, color_jitter_kwargs=None, random_affine_kwargs=None
+    ):
+        super().__init__()
+
+        def embed_crop(crop, augment=augment):
+            # crop is PIL.image of dtype="L" (so values range from 0 -> 255)
+            image = Image.new("L", (metadata.IMAGE_WIDTH, metadata.IMAGE_HEIGHT))
+
+            # Resize crop
+            crop_width, crop_height = crop.size
+            new_crop_height = metadata.IMAGE_HEIGHT
+            new_crop_width = int(new_crop_height * (crop_width / crop_height))
+            if augment:
+                # Add random stretching
+                new_crop_width = int(new_crop_width * random.uniform(0.9, 1.1))
+                new_crop_width = min(new_crop_width, metadata.IMAGE_WIDTH)
+            crop_resized = crop.resize(
+                (new_crop_width, new_crop_height), resample=Image.BILINEAR
+            )
+
+            # Embed in the image
+            x = min(metadata.CHAR_WIDTH, metadata.IMAGE_WIDTH - new_crop_width)
+            y = metadata.IMAGE_HEIGHT - new_crop_height
+
+            image.paste(crop_resized, (x, y))
+
+            return image
+
+        if color_jitter_kwargs is None:
+            color_jitter_kwargs = {"brightness": (0.8, 1.6)}
+        if random_affine_kwargs is None:
+            random_affine_kwargs = {
+                "degrees": 1,
+                "shear": (-30, 20),
+                "interpolation": T.InterpolationMode.BILINEAR,
+                "fill": 0,
+            }
+
+        pil_transform_list = [T.Lambda(embed_crop)]
+        if augment:
+            pil_transform_list += [
+                T.ColorJitter(**color_jitter_kwargs),
+                T.RandomAffine(**random_affine_kwargs),
+            ]
+        self.pil_transform = T.Compose(pil_transform_list)
diff --git a/text_recognizer/data/stems/paragraph.py b/text_recognizer/data/stems/paragraph.py
new file mode 100644
index 0000000..39e1e59
--- /dev/null
+++ b/text_recognizer/data/stems/paragraph.py
@@ -0,0 +1,66 @@
+"""Iam paragraph stem class."""
+import torchvision.transforms as T
+
+import text_recognizer.metadata.iam_paragraphs as metadata
+from text_recognizer.data.stems.image import ImageStem
+
+
+IMAGE_HEIGHT, IMAGE_WIDTH = metadata.IMAGE_HEIGHT, metadata.IMAGE_WIDTH
+IMAGE_SHAPE = metadata.IMAGE_SHAPE
+
+MAX_LABEL_LENGTH = metadata.MAX_LABEL_LENGTH
+
+
+class ParagraphStem(ImageStem):
+    """A stem for handling images that contain a paragraph of text."""
+
+    def __init__(
+        self,
+        augment=False,
+        color_jitter_kwargs=None,
+        random_affine_kwargs=None,
+        random_perspective_kwargs=None,
+        gaussian_blur_kwargs=None,
+        sharpness_kwargs=None,
+    ):
+        super().__init__()
+
+        if not augment:
+            self.pil_transform = T.Compose([T.CenterCrop(IMAGE_SHAPE)])
+        else:
+            if color_jitter_kwargs is None:
+                color_jitter_kwargs = {"brightness": 0.4, "contrast": 0.4}
+            if random_affine_kwargs is None:
+                random_affine_kwargs = {
+                    "degrees": 3,
+                    "shear": 6,
+                    "scale": (0.95, 1),
+                    "interpolation": T.InterpolationMode.BILINEAR,
+                }
+            if random_perspective_kwargs is None:
+                random_perspective_kwargs = {
+                    "distortion_scale": 0.2,
+                    "p": 0.5,
+                    "interpolation": T.InterpolationMode.BILINEAR,
+                }
+            if gaussian_blur_kwargs is None:
+                gaussian_blur_kwargs = {"kernel_size": (3, 3), "sigma": (0.1, 1.0)}
+            if sharpness_kwargs is None:
+                sharpness_kwargs = {"sharpness_factor": 2, "p": 0.5}
+
+            self.pil_transform = T.Compose(
+                [
+                    T.ColorJitter(**color_jitter_kwargs),
+                    T.RandomCrop(
+                        size=IMAGE_SHAPE,
+                        padding=None,
+                        pad_if_needed=True,
+                        fill=0,
+                        padding_mode="constant",
+                    ),
+                    T.RandomAffine(**random_affine_kwargs),
+                    T.RandomPerspective(**random_perspective_kwargs),
+                    T.GaussianBlur(**gaussian_blur_kwargs),
+                    T.RandomAdjustSharpness(**sharpness_kwargs),
+                ]
+            )