python-doctr 0.8.0__tar.gz → 0.8.1__tar.gz

{python-doctr-0.8.0/python_doctr.egg-info → python-doctr-0.8.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: python-doctr
-Version: 0.8.0
+Version: 0.8.1
 Summary: Document Text Recognition (docTR): deep Learning for high-performance OCR on documents.
 Author-email: Mindee <contact@mindee.com>
 Maintainer: François-Guillaume Fernandez, Charles Gaillard, Olivier Dulcy, Felix Dittrich
@@ -209,7 +209,7 @@ License:                                  Apache License
 Project-URL: documentation, https://mindee.github.io/doctr
 Project-URL: repository, https://github.com/mindee/doctr
 Project-URL: tracker, https://github.com/mindee/doctr/issues
-Project-URL: changelog, https://github.com/mindee/doctr/latest/changelog.html
+Project-URL: changelog, https://mindee.github.io/doctr/changelog.html
 Keywords: OCR,deep learning,computer vision,tensorflow,pytorch,text detection,text recognition
 Classifier: Development Status :: 4 - Beta
 Classifier: Intended Audience :: Developers
@@ -296,10 +296,10 @@ Requires-Dist: sphinx-tabs>=3.3.0; extra == "dev"
 Requires-Dist: furo>=2022.3.4; extra == "dev"
 
 <p align="center">
-  <img src="docs/images/Logo_doctr.gif" width="40%">
+  <img src="https://github.com/mindee/doctr/raw/main/docs/images/Logo_doctr.gif" width="40%">
 </p>
 
-[![Slack Icon](https://img.shields.io/badge/Slack-Community-4A154B?style=flat-square&logo=slack&logoColor=white)](https://slack.mindee.com) [![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](LICENSE) ![Build Status](https://github.com/mindee/doctr/workflows/builds/badge.svg) [![Docker Images](https://img.shields.io/badge/Docker-4287f5?style=flat&logo=docker&logoColor=white)](https://github.com/mindee/doctr/pkgs/container/doctr) [![codecov](https://codecov.io/gh/mindee/doctr/branch/main/graph/badge.svg?token=577MO567NM)](https://codecov.io/gh/mindee/doctr) [![CodeFactor](https://www.codefactor.io/repository/github/mindee/doctr/badge?s=bae07db86bb079ce9d6542315b8c6e70fa708a7e)](https://www.codefactor.io/repository/github/mindee/doctr) [![Codacy Badge](https://api.codacy.com/project/badge/Grade/340a76749b634586a498e1c0ab998f08)](https://app.codacy.com/gh/mindee/doctr?utm_source=github.com&utm_medium=referral&utm_content=mindee/doctr&utm_campaign=Badge_Grade) [![Doc Status](https://github.com/mindee/doctr/workflows/doc-status/badge.svg)](https://mindee.github.io/doctr) [![Pypi](https://img.shields.io/badge/pypi-v0.7.0-blue.svg)](https://pypi.org/project/python-doctr/) [![Hugging Face Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue)](https://huggingface.co/spaces/mindee/doctr) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/mindee/notebooks/blob/main/doctr/quicktour.ipynb)
+[![Slack Icon](https://img.shields.io/badge/Slack-Community-4A154B?style=flat-square&logo=slack&logoColor=white)](https://slack.mindee.com) [![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](LICENSE) ![Build Status](https://github.com/mindee/doctr/workflows/builds/badge.svg) [![Docker Images](https://img.shields.io/badge/Docker-4287f5?style=flat&logo=docker&logoColor=white)](https://github.com/mindee/doctr/pkgs/container/doctr) [![codecov](https://codecov.io/gh/mindee/doctr/branch/main/graph/badge.svg?token=577MO567NM)](https://codecov.io/gh/mindee/doctr) [![CodeFactor](https://www.codefactor.io/repository/github/mindee/doctr/badge?s=bae07db86bb079ce9d6542315b8c6e70fa708a7e)](https://www.codefactor.io/repository/github/mindee/doctr) [![Codacy Badge](https://api.codacy.com/project/badge/Grade/340a76749b634586a498e1c0ab998f08)](https://app.codacy.com/gh/mindee/doctr?utm_source=github.com&utm_medium=referral&utm_content=mindee/doctr&utm_campaign=Badge_Grade) [![Doc Status](https://github.com/mindee/doctr/workflows/doc-status/badge.svg)](https://mindee.github.io/doctr) [![Pypi](https://img.shields.io/badge/pypi-v0.8.0-blue.svg)](https://pypi.org/project/python-doctr/) [![Hugging Face Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue)](https://huggingface.co/spaces/mindee/doctr) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/mindee/notebooks/blob/main/doctr/quicktour.ipynb)
 
 
 **Optical Character Recognition made seamless & accessible to anyone, powered by TensorFlow 2 & PyTorch**
@@ -309,7 +309,7 @@ What you can expect from this repository:
 - efficient ways to parse textual information (localize and identify each word) from your documents
 - guidance on how to integrate this in your current architecture
 
-![OCR_example](docs/images/ocr.png)
+![OCR_example](https://github.com/mindee/doctr/raw/main/docs/images/ocr.png)
 
 ## Quick Tour
 
@@ -375,7 +375,7 @@ To interpret your model's predictions, you can visualize them interactively as f
 result.show()
 ```
 
-![Visualization sample](docs/images/doctr_example_script.gif)
+![Visualization sample](https://github.com/mindee/doctr/raw/main/docs/images/doctr_example_script.gif)
 
 Or even rebuild the original document from its predictions:
 
@@ -386,7 +386,7 @@ synthetic_pages = result.synthesize()
 plt.imshow(synthetic_pages[0]); plt.axis('off'); plt.show()
 ```
 
-![Synthesis sample](docs/images/synthesized_sample.png)
+![Synthesis sample](https://github.com/mindee/doctr/raw/main/docs/images/synthesized_sample.png)
 
 The `ocr_predictor` returns a `Document` object with a nested structure (with `Page`, `Block`, `Line`, `Word`, `Artefact`).
 To get a better understanding of our document model, check our [documentation](https://mindee.github.io/doctr/modules/io.html#document-structure):
@@ -425,7 +425,7 @@ The KIE predictor results per page are in a dictionary format with each key repr
 
 ### If you are looking for support from the Mindee team
 
-[![Bad OCR test detection image asking the developer if they need help](docs/images/doctr-need-help.png)](https://mindee.com/product/doctr)
+[![Bad OCR test detection image asking the developer if they need help](https://github.com/mindee/doctr/raw/main/docs/images/doctr-need-help.png)](https://mindee.com/product/doctr)
 
 ## Installation
 
@@ -494,6 +494,7 @@ Credits where it's due: this repository is implementing, among others, architect
 
 - DBNet: [Real-time Scene Text Detection with Differentiable Binarization](https://arxiv.org/pdf/1911.08947.pdf).
 - LinkNet: [LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation](https://arxiv.org/pdf/1707.03718.pdf)
+- FAST: [FAST: Faster Arbitrarily-Shaped Text Detector with Minimalist Kernel Representation](https://arxiv.org/pdf/2111.02394.pdf)
 
 ### Text Recognition
 
@@ -513,7 +514,7 @@ The full package documentation is available [here](https://mindee.github.io/doct
 
 A minimal demo app is provided for you to play with our end-to-end OCR models!
 
-![Demo app](docs/images/demo_update.png)
+![Demo app](https://github.com/mindee/doctr/raw/main/docs/images/demo_update.png)
 
 #### Live demo
 
@@ -553,11 +554,11 @@ USE_TORCH=1 streamlit run demo/app.py
 Instead of having your demo actually running Python, you would prefer to run everything in your web browser?
 Check out our [TensorFlow.js demo](https://github.com/mindee/doctr-tfjs-demo) to get started!
 
-![TFJS demo](docs/images/demo_illustration_mini.png)
+![TFJS demo](https://github.com/mindee/doctr/raw/main/docs/images/demo_illustration_mini.png)
 
 ### Docker container
 
-[We offers Docker container support for easy testing and deployment](https://github.com/mindee/doctr/packages).
+[We offer Docker container support for easy testing and deployment](https://github.com/mindee/doctr/pkgs/container/doctr).
 
 #### Using GPU with docTR Docker Images
 
@@ -673,8 +674,8 @@ If you wish to cite this project, feel free to use this [BibTeX](http://www.bibt
 
 If you scrolled down to this section, you most likely appreciate open source. Do you feel like extending the range of our supported characters? Or perhaps submitting a paper implementation? Or contributing in any other way?
 
-You're in luck, we compiled a short guide (cf. [`CONTRIBUTING`](CONTRIBUTING.md)) for you to easily do so!
+You're in luck, we compiled a short guide (cf. [`CONTRIBUTING`](https://mindee.github.io/doctr/contributing/contributing.html)) for you to easily do so!
 
 ## License
 
-Distributed under the Apache 2.0 License. See [`LICENSE`](LICENSE) for more information.
+Distributed under the Apache 2.0 License. See [`LICENSE`](https://github.com/mindee/doctr?tab=Apache-2.0-1-ov-file#readme) for more information.

{python-doctr-0.8.0 → python-doctr-0.8.1}/README.md

@@ -1,8 +1,8 @@
 <p align="center">
-  <img src="docs/images/Logo_doctr.gif" width="40%">
+  <img src="https://github.com/mindee/doctr/raw/main/docs/images/Logo_doctr.gif" width="40%">
 </p>
 
-[![Slack Icon](https://img.shields.io/badge/Slack-Community-4A154B?style=flat-square&logo=slack&logoColor=white)](https://slack.mindee.com) [![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](LICENSE) ![Build Status](https://github.com/mindee/doctr/workflows/builds/badge.svg) [![Docker Images](https://img.shields.io/badge/Docker-4287f5?style=flat&logo=docker&logoColor=white)](https://github.com/mindee/doctr/pkgs/container/doctr) [![codecov](https://codecov.io/gh/mindee/doctr/branch/main/graph/badge.svg?token=577MO567NM)](https://codecov.io/gh/mindee/doctr) [![CodeFactor](https://www.codefactor.io/repository/github/mindee/doctr/badge?s=bae07db86bb079ce9d6542315b8c6e70fa708a7e)](https://www.codefactor.io/repository/github/mindee/doctr) [![Codacy Badge](https://api.codacy.com/project/badge/Grade/340a76749b634586a498e1c0ab998f08)](https://app.codacy.com/gh/mindee/doctr?utm_source=github.com&utm_medium=referral&utm_content=mindee/doctr&utm_campaign=Badge_Grade) [![Doc Status](https://github.com/mindee/doctr/workflows/doc-status/badge.svg)](https://mindee.github.io/doctr) [![Pypi](https://img.shields.io/badge/pypi-v0.7.0-blue.svg)](https://pypi.org/project/python-doctr/) [![Hugging Face Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue)](https://huggingface.co/spaces/mindee/doctr) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/mindee/notebooks/blob/main/doctr/quicktour.ipynb)
+[![Slack Icon](https://img.shields.io/badge/Slack-Community-4A154B?style=flat-square&logo=slack&logoColor=white)](https://slack.mindee.com) [![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](LICENSE) ![Build Status](https://github.com/mindee/doctr/workflows/builds/badge.svg) [![Docker Images](https://img.shields.io/badge/Docker-4287f5?style=flat&logo=docker&logoColor=white)](https://github.com/mindee/doctr/pkgs/container/doctr) [![codecov](https://codecov.io/gh/mindee/doctr/branch/main/graph/badge.svg?token=577MO567NM)](https://codecov.io/gh/mindee/doctr) [![CodeFactor](https://www.codefactor.io/repository/github/mindee/doctr/badge?s=bae07db86bb079ce9d6542315b8c6e70fa708a7e)](https://www.codefactor.io/repository/github/mindee/doctr) [![Codacy Badge](https://api.codacy.com/project/badge/Grade/340a76749b634586a498e1c0ab998f08)](https://app.codacy.com/gh/mindee/doctr?utm_source=github.com&utm_medium=referral&utm_content=mindee/doctr&utm_campaign=Badge_Grade) [![Doc Status](https://github.com/mindee/doctr/workflows/doc-status/badge.svg)](https://mindee.github.io/doctr) [![Pypi](https://img.shields.io/badge/pypi-v0.8.0-blue.svg)](https://pypi.org/project/python-doctr/) [![Hugging Face Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue)](https://huggingface.co/spaces/mindee/doctr) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/mindee/notebooks/blob/main/doctr/quicktour.ipynb)
 
 
 **Optical Character Recognition made seamless & accessible to anyone, powered by TensorFlow 2 & PyTorch**
@@ -12,7 +12,7 @@ What you can expect from this repository:
 - efficient ways to parse textual information (localize and identify each word) from your documents
 - guidance on how to integrate this in your current architecture
 
-![OCR_example](docs/images/ocr.png)
+![OCR_example](https://github.com/mindee/doctr/raw/main/docs/images/ocr.png)
 
 ## Quick Tour
 
@@ -78,7 +78,7 @@ To interpret your model's predictions, you can visualize them interactively as f
 result.show()
 ```
 
-![Visualization sample](docs/images/doctr_example_script.gif)
+![Visualization sample](https://github.com/mindee/doctr/raw/main/docs/images/doctr_example_script.gif)
 
 Or even rebuild the original document from its predictions:
 
@@ -89,7 +89,7 @@ synthetic_pages = result.synthesize()
 plt.imshow(synthetic_pages[0]); plt.axis('off'); plt.show()
 ```
 
-![Synthesis sample](docs/images/synthesized_sample.png)
+![Synthesis sample](https://github.com/mindee/doctr/raw/main/docs/images/synthesized_sample.png)
 
 The `ocr_predictor` returns a `Document` object with a nested structure (with `Page`, `Block`, `Line`, `Word`, `Artefact`).
 To get a better understanding of our document model, check our [documentation](https://mindee.github.io/doctr/modules/io.html#document-structure):
@@ -128,7 +128,7 @@ The KIE predictor results per page are in a dictionary format with each key repr
 
 ### If you are looking for support from the Mindee team
 
-[![Bad OCR test detection image asking the developer if they need help](docs/images/doctr-need-help.png)](https://mindee.com/product/doctr)
+[![Bad OCR test detection image asking the developer if they need help](https://github.com/mindee/doctr/raw/main/docs/images/doctr-need-help.png)](https://mindee.com/product/doctr)
 
 ## Installation
 
@@ -197,6 +197,7 @@ Credits where it's due: this repository is implementing, among others, architect
 
 - DBNet: [Real-time Scene Text Detection with Differentiable Binarization](https://arxiv.org/pdf/1911.08947.pdf).
 - LinkNet: [LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation](https://arxiv.org/pdf/1707.03718.pdf)
+- FAST: [FAST: Faster Arbitrarily-Shaped Text Detector with Minimalist Kernel Representation](https://arxiv.org/pdf/2111.02394.pdf)
 
 ### Text Recognition
 
@@ -216,7 +217,7 @@ The full package documentation is available [here](https://mindee.github.io/doct
 
 A minimal demo app is provided for you to play with our end-to-end OCR models!
 
-![Demo app](docs/images/demo_update.png)
+![Demo app](https://github.com/mindee/doctr/raw/main/docs/images/demo_update.png)
 
 #### Live demo
 
@@ -256,11 +257,11 @@ USE_TORCH=1 streamlit run demo/app.py
 Instead of having your demo actually running Python, you would prefer to run everything in your web browser?
 Check out our [TensorFlow.js demo](https://github.com/mindee/doctr-tfjs-demo) to get started!
 
-![TFJS demo](docs/images/demo_illustration_mini.png)
+![TFJS demo](https://github.com/mindee/doctr/raw/main/docs/images/demo_illustration_mini.png)
 
 ### Docker container
 
-[We offers Docker container support for easy testing and deployment](https://github.com/mindee/doctr/packages).
+[We offer Docker container support for easy testing and deployment](https://github.com/mindee/doctr/pkgs/container/doctr).
 
 #### Using GPU with docTR Docker Images
 
@@ -376,8 +377,8 @@ If you wish to cite this project, feel free to use this [BibTeX](http://www.bibt
 
 If you scrolled down to this section, you most likely appreciate open source. Do you feel like extending the range of our supported characters? Or perhaps submitting a paper implementation? Or contributing in any other way?
 
-You're in luck, we compiled a short guide (cf. [`CONTRIBUTING`](CONTRIBUTING.md)) for you to easily do so!
+You're in luck, we compiled a short guide (cf. [`CONTRIBUTING`](https://mindee.github.io/doctr/contributing/contributing.html)) for you to easily do so!
 
 ## License
 
-Distributed under the Apache 2.0 License. See [`LICENSE`](LICENSE) for more information.
+Distributed under the Apache 2.0 License. See [`LICENSE`](https://github.com/mindee/doctr?tab=Apache-2.0-1-ov-file#readme) for more information.

{python-doctr-0.8.0 → python-doctr-0.8.1}/doctr/models/detection/__init__.py

@@ -1,3 +1,4 @@
 from .differentiable_binarization import *
 from .linknet import *
+from .fast import *
 from .zoo import *

{python-doctr-0.8.0 → python-doctr-0.8.1}/doctr/models/detection/differentiable_binarization/tensorflow.py

@@ -147,24 +147,20 @@ class DBNet(_DBNet, keras.Model, NestedObject):
         _inputs = [layers.Input(shape=in_shape[1:]) for in_shape in self.feat_extractor.output_shape]
         output_shape = tuple(self.fpn(_inputs).shape)
 
-        self.probability_head = keras.Sequential(
-            [
-                *conv_sequence(64, "relu", True, kernel_size=3, input_shape=output_shape[1:]),
-                layers.Conv2DTranspose(64, 2, strides=2, use_bias=False, kernel_initializer="he_normal"),
-                layers.BatchNormalization(),
-                layers.Activation("relu"),
-                layers.Conv2DTranspose(num_classes, 2, strides=2, kernel_initializer="he_normal"),
-            ]
-        )
-        self.threshold_head = keras.Sequential(
-            [
-                *conv_sequence(64, "relu", True, kernel_size=3, input_shape=output_shape[1:]),
-                layers.Conv2DTranspose(64, 2, strides=2, use_bias=False, kernel_initializer="he_normal"),
-                layers.BatchNormalization(),
-                layers.Activation("relu"),
-                layers.Conv2DTranspose(num_classes, 2, strides=2, kernel_initializer="he_normal"),
-            ]
-        )
+        self.probability_head = keras.Sequential([
+            *conv_sequence(64, "relu", True, kernel_size=3, input_shape=output_shape[1:]),
+            layers.Conv2DTranspose(64, 2, strides=2, use_bias=False, kernel_initializer="he_normal"),
+            layers.BatchNormalization(),
+            layers.Activation("relu"),
+            layers.Conv2DTranspose(num_classes, 2, strides=2, kernel_initializer="he_normal"),
+        ])
+        self.threshold_head = keras.Sequential([
+            *conv_sequence(64, "relu", True, kernel_size=3, input_shape=output_shape[1:]),
+            layers.Conv2DTranspose(64, 2, strides=2, use_bias=False, kernel_initializer="he_normal"),
+            layers.BatchNormalization(),
+            layers.Activation("relu"),
+            layers.Conv2DTranspose(num_classes, 2, strides=2, kernel_initializer="he_normal"),
+        ])
 
         self.postprocessor = DBPostProcessor(
             assume_straight_pages=assume_straight_pages, bin_thresh=bin_thresh, box_thresh=box_thresh

python-doctr-0.8.1/doctr/models/detection/fast/base.py

@@ -0,0 +1,256 @@
+# Copyright (C) 2021-2024, Mindee.
+
+# This program is licensed under the Apache License 2.0.
+# See LICENSE or go to <https://opensource.org/licenses/Apache-2.0> for full license details.
+
+# Credits: post-processing adapted from https://github.com/xuannianz/DifferentiableBinarization
+
+from typing import Dict, List, Tuple, Union
+
+import cv2
+import numpy as np
+import pyclipper
+from shapely.geometry import Polygon
+
+from doctr.models.core import BaseModel
+
+from ..core import DetectionPostProcessor
+
+__all__ = ["_FAST", "FASTPostProcessor"]
+
+
+class FASTPostProcessor(DetectionPostProcessor):
+    """Implements a post processor for FAST model.
+
+    Args:
+    ----
+        bin_thresh: threshold used to binzarized p_map at inference time
+        box_thresh: minimal objectness score to consider a box
+        assume_straight_pages: whether the inputs were expected to have horizontal text elements
+    """
+
+    def __init__(
+        self,
+        bin_thresh: float = 0.3,
+        box_thresh: float = 0.1,
+        assume_straight_pages: bool = True,
+    ) -> None:
+        super().__init__(box_thresh, bin_thresh, assume_straight_pages)
+        self.unclip_ratio = 1.0
+
+    def polygon_to_box(
+        self,
+        points: np.ndarray,
+    ) -> np.ndarray:
+        """Expand a polygon (points) by a factor unclip_ratio, and returns a polygon
+
+        Args:
+        ----
+            points: The first parameter.
+
+        Returns:
+        -------
+            a box in absolute coordinates (xmin, ymin, xmax, ymax) or (4, 2) array (quadrangle)
+        """
+        if not self.assume_straight_pages:
+            # Compute the rectangle polygon enclosing the raw polygon
+            rect = cv2.minAreaRect(points)
+            points = cv2.boxPoints(rect)
+            # Add 1 pixel to correct cv2 approx
+            area = (rect[1][0] + 1) * (1 + rect[1][1])
+            length = 2 * (rect[1][0] + rect[1][1]) + 2
+        else:
+            poly = Polygon(points)
+            area = poly.area
+            length = poly.length
+        distance = area * self.unclip_ratio / length  # compute distance to expand polygon
+        offset = pyclipper.PyclipperOffset()
+        offset.AddPath(points, pyclipper.JT_ROUND, pyclipper.ET_CLOSEDPOLYGON)
+        _points = offset.Execute(distance)
+        # Take biggest stack of points
+        idx = 0
+        if len(_points) > 1:
+            max_size = 0
+            for _idx, p in enumerate(_points):
+                if len(p) > max_size:
+                    idx = _idx
+                    max_size = len(p)
+            # We ensure that _points can be correctly casted to a ndarray
+            _points = [_points[idx]]
+        expanded_points: np.ndarray = np.asarray(_points)  # expand polygon
+        if len(expanded_points) < 1:
+            return None  # type: ignore[return-value]
+        return (
+            cv2.boundingRect(expanded_points)  # type: ignore[return-value]
+            if self.assume_straight_pages
+            else np.roll(cv2.boxPoints(cv2.minAreaRect(expanded_points)), -1, axis=0)
+        )
+
+    def bitmap_to_boxes(
+        self,
+        pred: np.ndarray,
+        bitmap: np.ndarray,
+    ) -> np.ndarray:
+        """Compute boxes from a bitmap/pred_map: find connected components then filter boxes
+
+        Args:
+        ----
+            pred: Pred map from differentiable linknet output
+            bitmap: Bitmap map computed from pred (binarized)
+            angle_tol: Comparison tolerance of the angle with the median angle across the page
+            ratio_tol: Under this limit aspect ratio, we cannot resolve the direction of the crop
+
+        Returns:
+        -------
+            np tensor boxes for the bitmap, each box is a 6-element list
+                containing x, y, w, h, alpha, score for the box
+        """
+        height, width = bitmap.shape[:2]
+        boxes: List[Union[np.ndarray, List[float]]] = []
+        # get contours from connected components on the bitmap
+        contours, _ = cv2.findContours(bitmap.astype(np.uint8), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+        for contour in contours:
+            # Check whether smallest enclosing bounding box is not too small
+            if np.any(contour[:, 0].max(axis=0) - contour[:, 0].min(axis=0) < 2):
+                continue
+            # Compute objectness
+            if self.assume_straight_pages:
+                x, y, w, h = cv2.boundingRect(contour)
+                points: np.ndarray = np.array([[x, y], [x, y + h], [x + w, y + h], [x + w, y]])
+                score = self.box_score(pred, points, assume_straight_pages=True)
+            else:
+                score = self.box_score(pred, contour, assume_straight_pages=False)
+
+            if score < self.box_thresh:  # remove polygons with a weak objectness
+                continue
+
+            if self.assume_straight_pages:
+                _box = self.polygon_to_box(points)
+            else:
+                _box = self.polygon_to_box(np.squeeze(contour))
+
+            if self.assume_straight_pages:
+                # compute relative polygon to get rid of img shape
+                x, y, w, h = _box
+                xmin, ymin, xmax, ymax = x / width, y / height, (x + w) / width, (y + h) / height
+                boxes.append([xmin, ymin, xmax, ymax, score])
+            else:
+                # compute relative box to get rid of img shape
+                _box[:, 0] /= width
+                _box[:, 1] /= height
+                boxes.append(_box)
+
+        if not self.assume_straight_pages:
+            return np.clip(np.asarray(boxes), 0, 1) if len(boxes) > 0 else np.zeros((0, 4, 2), dtype=pred.dtype)
+        else:
+            return np.clip(np.asarray(boxes), 0, 1) if len(boxes) > 0 else np.zeros((0, 5), dtype=pred.dtype)
+
+
+class _FAST(BaseModel):
+    """FAST as described in `"FAST: Faster Arbitrarily-Shaped Text Detector with Minimalist Kernel Representation"
+    <https://arxiv.org/pdf/2111.02394.pdf>`_.
+    """
+
+    min_size_box: int = 3
+    assume_straight_pages: bool = True
+    shrink_ratio = 0.1
+
+    def build_target(
+        self,
+        target: List[Dict[str, np.ndarray]],
+        output_shape: Tuple[int, int, int],
+        channels_last: bool = True,
+    ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """Build the target, and it's mask to be used from loss computation.
+
+        Args:
+        ----
+            target: target coming from dataset
+            output_shape: shape of the output of the model without batch_size
+            channels_last: whether channels are last or not
+
+        Returns:
+        -------
+            the new formatted target, mask and shrunken text kernel
+        """
+        if any(t.dtype != np.float32 for tgt in target for t in tgt.values()):
+            raise AssertionError("the expected dtype of target 'boxes' entry is 'np.float32'.")
+        if any(np.any((t[:, :4] > 1) | (t[:, :4] < 0)) for tgt in target for t in tgt.values()):
+            raise ValueError("the 'boxes' entry of the target is expected to take values between 0 & 1.")
+
+        h: int
+        w: int
+        if channels_last:
+            h, w, num_classes = output_shape
+        else:
+            num_classes, h, w = output_shape
+        target_shape = (len(target), num_classes, h, w)
+
+        seg_target: np.ndarray = np.zeros(target_shape, dtype=np.uint8)
+        seg_mask: np.ndarray = np.ones(target_shape, dtype=bool)
+        shrunken_kernel: np.ndarray = np.zeros(target_shape, dtype=np.uint8)
+
+        for idx, tgt in enumerate(target):
+            for class_idx, _tgt in enumerate(tgt.values()):
+                # Draw each polygon on gt
+                if _tgt.shape[0] == 0:
+                    # Empty image, full masked
+                    seg_mask[idx, class_idx] = False
+
+                # Absolute bounding boxes
+                abs_boxes = _tgt.copy()
+
+                if abs_boxes.ndim == 3:
+                    abs_boxes[:, :, 0] *= w
+                    abs_boxes[:, :, 1] *= h
+                    polys = abs_boxes
+                    boxes_size = np.linalg.norm(abs_boxes[:, 2, :] - abs_boxes[:, 0, :], axis=-1)
+                    abs_boxes = np.concatenate((abs_boxes.min(1), abs_boxes.max(1)), -1).round().astype(np.int32)
+                else:
+                    abs_boxes[:, [0, 2]] *= w
+                    abs_boxes[:, [1, 3]] *= h
+                    abs_boxes = abs_boxes.round().astype(np.int32)
+                    polys = np.stack(
+                        [
+                            abs_boxes[:, [0, 1]],
+                            abs_boxes[:, [0, 3]],
+                            abs_boxes[:, [2, 3]],
+                            abs_boxes[:, [2, 1]],
+                        ],
+                        axis=1,
+                    )
+                    boxes_size = np.minimum(abs_boxes[:, 2] - abs_boxes[:, 0], abs_boxes[:, 3] - abs_boxes[:, 1])
+
+                for poly, box, box_size in zip(polys, abs_boxes, boxes_size):
+                    # Mask boxes that are too small
+                    if box_size < self.min_size_box:
+                        seg_mask[idx, class_idx, box[1] : box[3] + 1, box[0] : box[2] + 1] = False
+                        continue
+
+                    # Negative shrink for gt, as described in paper
+                    polygon = Polygon(poly)
+                    distance = polygon.area * (1 - np.power(self.shrink_ratio, 2)) / polygon.length
+                    subject = [tuple(coor) for coor in poly]
+                    padding = pyclipper.PyclipperOffset()
+                    padding.AddPath(subject, pyclipper.JT_ROUND, pyclipper.ET_CLOSEDPOLYGON)
+                    shrunken = padding.Execute(-distance)
+
+                    # Draw polygon on gt if it is valid
+                    if len(shrunken) == 0:
+                        seg_mask[idx, class_idx, box[1] : box[3] + 1, box[0] : box[2] + 1] = False
+                        continue
+                    shrunken = np.array(shrunken[0]).reshape(-1, 2)
+                    if shrunken.shape[0] <= 2 or not Polygon(shrunken).is_valid:
+                        seg_mask[idx, class_idx, box[1] : box[3] + 1, box[0] : box[2] + 1] = False
+                        continue
+                    cv2.fillPoly(shrunken_kernel[idx, class_idx], [shrunken.astype(np.int32)], 1.0)  # type: ignore[call-overload]
+                    # draw the original polygon on the segmentation target
+                    cv2.fillPoly(seg_target[idx, class_idx], [poly.astype(np.int32)], 1.0)  # type: ignore[call-overload]
+
+        # Don't forget to switch back to channel last if Tensorflow is used
+        if channels_last:
+            seg_target = seg_target.transpose((0, 2, 3, 1))
+            seg_mask = seg_mask.transpose((0, 2, 3, 1))
+            shrunken_kernel = shrunken_kernel.transpose((0, 2, 3, 1))
+
+        return seg_target, seg_mask, shrunken_kernel