|
| 1 | +--- |
| 2 | +layout: blog_detail |
| 3 | +title: "docTR joins PyTorch Ecosystem: From Pixels to Data, Building a Recognition Pipeline with PyTorch and docTR" |
| 4 | +author: Olivier Dulcy & Sebastian Olivera, Mindee |
| 5 | +--- |
| 6 | + |
| 7 | +{:style="width:100%;display: block;max-width:400px; margin-left:auto; margin-right:auto;"} |
| 8 | + |
| 9 | +We’re thrilled to announce that the docTR project has been integrated into the PyTorch ecosystem! This integration ensures that docTR aligns with PyTorch’s standards and practices, giving developers a reliable, community-backed solution for powerful OCR workflows. |
| 10 | + |
| 11 | +**For more information on what it means to be a PyTorch ecosystem project, see the [PyTorch Ecosystem Tools page](https://pytorch.org/ecosystem/).** |
| 12 | + |
| 13 | + |
| 14 | +## About docTR |
| 15 | + |
| 16 | +docTR is an Apache 2.0 project developed and distributed by [Mindee](https://www.mindee.com/) to help developers integrate OCR capabilities into applications with no prior knowledge required. |
| 17 | + |
| 18 | +To quickly and efficiently extract text information, docTR uses a two-stage approach: |
| 19 | + |
| 20 | + |
| 21 | + |
| 22 | +* First, it performs text **detection** to localize words. |
| 23 | +* Then, it conducts text **recognition** to identify all characters in a word. |
| 24 | + |
| 25 | +**Detection** and **recognition** are performed by state-of-the-art models written in PyTorch. To learn more about this approach, you can refer [to the docTR documentation](https://mindee.github.io/doctr/using_doctr/using_models.html). |
| 26 | + |
| 27 | +docTR enhances the user experience in PyTorch projects by providing high-performance OCR capabilities right out of the box. Its specially designed models require minimal to no fine-tuning for common use cases, allowing developers to quickly integrate advanced document analysis features. |
| 28 | + |
| 29 | + |
| 30 | +## Local installation |
| 31 | + |
| 32 | +docTR requires Python >= 3.10 and supports Windows, Mac and Linux. Please refer to our [README](https://github.com/mindee/doctr?tab=readme-ov-file#installation) for necessary dependencies for MacBook with the M1 chip. |
| 33 | + |
| 34 | +``` |
| 35 | +pip3 install -U pip |
| 36 | +pip3 install "python-doctr[torch,viz]" |
| 37 | +``` |
| 38 | + |
| 39 | +This will install docTR along with the latest version of PyTorch. |
| 40 | + |
| 41 | + |
| 42 | +``` |
| 43 | +Note: docTR also provides docker images for an easy deployment, such as a part of Kubernetes cluster. |
| 44 | +``` |
| 45 | + |
| 46 | + |
| 47 | + |
| 48 | +## Text recognition |
| 49 | + |
| 50 | +Now, let’s try docTR’s OCR recognition on this sample: |
| 51 | + |
| 52 | + |
| 53 | +{:style="width:100%;display: block;max-width:300px; margin-left:auto; margin-right:auto;"} |
| 54 | + |
| 55 | + |
| 56 | +The OCR recognition model expects an image with only one word on it and will output the predicted word with a confidence score. You can use the following snippet to test OCR capabilities from docTR: |
| 57 | + |
| 58 | +``` |
| 59 | +python |
| 60 | +from doctr.io import DocumentFile |
| 61 | +from doctr.models import recognition_predictor |
| 62 | +
|
| 63 | +doc = DocumentFile.from_images("/path/to/image") |
| 64 | +
|
| 65 | +# Load the OCR model |
| 66 | +# This will download pre-trained models hosted by Mindee |
| 67 | +model = recognition_predictor(pretrained=True) |
| 68 | +
|
| 69 | +result = model(doc) |
| 70 | +print(result) |
| 71 | +``` |
| 72 | + |
| 73 | +Here, the most important line of code is `model = recognition_predictor(pretrained=True)`. This will load a default text recognition model, `crnn_vgg16_bn`, but you can select other models through the `arch` parameter. You can check out the [available architectures](https://mindee.github.io/doctr/using_doctr/using_models.html). |
| 74 | + |
| 75 | +When run on the sample, the recognition predictor retrieves the following data: `[('MAGAZINE', 0.9872216582298279)]` |
| 76 | + |
| 77 | + |
| 78 | +``` |
| 79 | +Note: using the DocumentFile object docTR provides an easy way to manipulate PDF or Images. |
| 80 | +``` |
| 81 | + |
| 82 | + |
| 83 | + |
| 84 | +## Text detection |
| 85 | + |
| 86 | +The last example was a crop on a single word. Now, what about an image with several words on it, like this one? |
| 87 | + |
| 88 | + |
| 89 | +{:style="width:100%;display: block;max-width:300px; margin-left:auto; margin-right:auto;"} |
| 90 | + |
| 91 | + |
| 92 | +A text detection model is used before the text recognition to output a segmentation map representing the location of the text. Following that, the text recognition is applied on every detected patch. |
| 93 | + |
| 94 | +Below is a snippet to run only the detection part: |
| 95 | + |
| 96 | +``` |
| 97 | +from doctr.io import DocumentFile |
| 98 | +from doctr.models import detection_predictor |
| 99 | +from matplotlib import pyplot as plt |
| 100 | +from doctr.utils.geometry import detach_scores |
| 101 | +from doctr.utils.visualization import draw_boxes |
| 102 | +
|
| 103 | +doc = DocumentFile.from_images("path/to/my/file") |
| 104 | +model = detection_predictor(pretrained=True) |
| 105 | +
|
| 106 | +result = model(doc) |
| 107 | +
|
| 108 | +draw_boxes(detach_scores([result[0]["words"]])[0][0], doc[0]) |
| 109 | +plt.axis('off') |
| 110 | +plt.show() |
| 111 | +``` |
| 112 | + |
| 113 | +Running it on the full sample yields the following: |
| 114 | + |
| 115 | + |
| 116 | +{:style="width:100%;display: block;max-width:300px; margin-left:auto; margin-right:auto;"} |
| 117 | + |
| 118 | + |
| 119 | +Similarly to the text recognition, `detection_predictor` will load a default model (`fast_base` here). You can also load another one by providing it through the `arch` parameter. |
| 120 | + |
| 121 | + |
| 122 | +## The full implementation |
| 123 | + |
| 124 | +Now, let’s plug both components into the same pipeline. |
| 125 | + |
| 126 | +Conveniently, docTR provides a wrapper that does exactly that for us: |
| 127 | + |
| 128 | +``` |
| 129 | +from doctr.io import DocumentFile |
| 130 | +from doctr.models import ocr_predictor |
| 131 | +
|
| 132 | +doc = DocumentFile.from_images("/path/to/image") |
| 133 | +
|
| 134 | +model = ocr_predictor(pretrained=True, assume_straight_pages=False) |
| 135 | +
|
| 136 | +result = model(doc) |
| 137 | +result.show() |
| 138 | +``` |
| 139 | + |
| 140 | +{:style="width:100%;display: block;max-width:300px; margin-left:auto; margin-right:auto;"} |
| 141 | + |
| 142 | +The last line should display a matplotlib window which shows the detected patches. Hovering the mouse over them will display their contents. |
| 143 | + |
| 144 | +You can also do more with this output, such as reconstituting a synthetic document like so: |
| 145 | + |
| 146 | +``` |
| 147 | +import matplotlib.pyplot as plt |
| 148 | +
|
| 149 | +synthetic_pages = result.synthesize() |
| 150 | +plt.imshow(synthetic_pages[0]) |
| 151 | +plt.axis('off') |
| 152 | +plt.show() |
| 153 | +``` |
| 154 | + |
| 155 | +{:style="width:100%;display: block;max-width:300px; margin-left:auto; margin-right:auto;"} |
| 156 | + |
| 157 | + |
| 158 | +The pipeline is highly customizable, where you can modify the detection or recognition model behaviors by passing arguments to the `ocr_predictor`. Please refer to the [documentation](https://mindee.github.io/doctr/using_doctr/using_models.html) to learn more about it. |
| 159 | + |
| 160 | + |
| 161 | +## Conclusion |
| 162 | + |
| 163 | +We’re excited to welcome docTR into the PyTorch Ecosystem, where it seamlessly integrates with PyTorch pipelines to deliver state-of-the-art OCR capabilities right out of the box. |
| 164 | + |
| 165 | +By empowering developers to quickly extract text from images or PDFs using familiar tooling, docTR simplifies complex document analysis tasks and enhances the overall PyTorch experience. |
| 166 | + |
| 167 | +We invite you to explore the [docTR GitHub repository ](https://github.com/mindee/doctr), join the [docTR community on Slack ](https://slack.mindee.com/), and reach out at [email protected] for inquiries or collaboration opportunities. |
| 168 | + |
| 169 | +Together, we can continue to push the boundaries of document understanding and develop even more powerful, accessible tools for everyone in the PyTorch community. |
0 commit comments