{
  "id": 80822,
  "title": "Images with text ocr",
  "url": "/competitions/humpback-whale-identification/discussion/80822",
  "author_name": "",
  "post_date": "2019-02-16T21:11:17.501820900Z",
  "votes": 22,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Following the excellent <a href=\"https://www.kaggle.com/voglinio/images-containing-text\">kernel</a> by <a href=\"https://www.kaggle.com/voglinio\">Costas Voglis</a> I tried a little ocr on the images containing text using pytesseract. Since I cannot figure out how to get tesseract to work on kaggle I am posting the results here. </p>\n\n<p>After some preprocessing it does a decent job with the darker text though on the images with fainter text and strange font its not very good. Overlap between train and test is minimal, I found less than 50 images with the same #xxxx between them, but the data are the data and we must use what we have. There are probably more matches but I only looked at the ones with the #. In any case I strongly recommend verifying everything yourself.\nGood fortunes! </p>",
  "messages": [
    {
      "id": "472894",
      "postDate": "02/16/2019 21:11:17",
      "content": "<p>Following the excellent <a href=\"https://www.kaggle.com/voglinio/images-containing-text\">kernel</a> by <a href=\"https://www.kaggle.com/voglinio\">Costas Voglis</a> I tried a little ocr on the images containing text using pytesseract. Since I cannot figure out how to get tesseract to work on kaggle I am posting the results here. </p>\n\n<p>After some preprocessing it does a decent job with the darker text though on the images with fainter text and strange font its not very good. Overlap between train and test is minimal, I found less than 50 images with the same #xxxx between them, but the data are the data and we must use what we have. There are probably more matches but I only looked at the ones with the #. In any case I strongly recommend verifying everything yourself.\nGood fortunes! </p>",
      "rawMarkdown": "Following the excellent [kernel][1] by [Costas Voglis][2] I tried a little ocr on the images containing text using pytesseract. Since I cannot figure out how to get tesseract to work on kaggle I am posting the results here. \n\nAfter some preprocessing it does a decent job with the darker text though on the images with fainter text and strange font its not very good. Overlap between train and test is minimal, I found less than 50 images with the same #xxxx between them, but the data are the data and we must use what we have. There are probably more matches but I only looked at the ones with the #. In any case I strongly recommend verifying everything yourself.\nGood fortunes! \n\n\n  [1]: https://www.kaggle.com/voglinio/images-containing-text\n  [2]: https://www.kaggle.com/voglinio",
      "votes": null
    },
    {
      "id": "474291",
      "postDate": "02/19/2019 07:07:30",
      "content": "<p>Thanks a lot! From your findings, I update 7 whales that my model does not correctly predicted.</p>",
      "rawMarkdown": "Thanks a lot! From your findings, I update 7 whales that my model does not correctly predicted.",
      "votes": null
    },
    {
      "id": "474539",
      "postDate": "02/19/2019 14:22:23",
      "content": "<p>instead of predicting the text, for those whale id with text, i augment their images with text. text is treated as addition visual appearance features. I also create new fake image = empty + text</p>\n\n<p>an example of image that is difficult to match with whale image, bur easy with text</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/474539/11356/mtach_by_text.png\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "instead of predicting the text, for those whale id with text, i augment their images with text. text is treated as addition visual appearance features. I also create new fake image = empty + text\n\nan example of image that is difficult to match with whale image, bur easy with text\n\n   ![enter image description here][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/474539/11356/mtach_by_text.png",
      "votes": null
    },
    {
      "id": "474764",
      "postDate": "02/19/2019 19:15:34",
      "content": "<p>That’s really cool. I thought about doing something like that, possibly by chopping up the text-detection network using it in the whale network but I don’t really know how to do it. </p>\n\n<p>As for the ocr, I know it’s not the sort of thing the sponsors want as it’s unlikely users of their site will upload new images they have taken with added text matching something in the database, but matching the text to get a few more training examples seemed reasonable. </p>",
      "rawMarkdown": "That’s really cool. I thought about doing something like that, possibly by chopping up the text-detection network using it in the whale network but I don’t really know how to do it. \n\nAs for the ocr, I know it’s not the sort of thing the sponsors want as it’s unlikely users of their site will upload new images they have taken with added text matching something in the database, but matching the text to get a few more training examples seemed reasonable.",
      "votes": null
    },
    {
      "id": "475680",
      "postDate": "02/21/2019 04:09:20",
      "content": "<p>This is brilliant! Thanks for sharing. </p>",
      "rawMarkdown": "This is brilliant! Thanks for sharing.",
      "votes": null
    },
    {
      "id": "1076117",
      "postDate": "11/12/2020 09:21:04",
      "content": "<p>I write a script for a beginner who wants to Learn OCR Step by Step. Just go and watch and support me if you like:-</p>\n<p><a href=\"https://www.kaggle.com/naim99/ocr-text-recognition-ocr-space-api-tesseract\" target=\"_blank\">https://www.kaggle.com/naim99/ocr-text-recognition-ocr-space-api-tesseract</a></p>",
      "rawMarkdown": "I write a script for a beginner who wants to Learn OCR Step by Step. Just go and watch and support me if you like:-\n\nhttps://www.kaggle.com/naim99/ocr-text-recognition-ocr-space-api-tesseract",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1076117,
      "author_name": "naim99",
      "author_url": "",
      "post_date": "11/12/2020 09:21:04",
      "content": "<p>I write a script for a beginner who wants to Learn OCR Step by Step. Just go and watch and support me if you like:-</p>\n<p><a href=\"https://www.kaggle.com/naim99/ocr-text-recognition-ocr-space-api-tesseract\" target=\"_blank\">https://www.kaggle.com/naim99/ocr-text-recognition-ocr-space-api-tesseract</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 474291,
      "author_name": "sophie0308",
      "author_url": "",
      "post_date": "02/19/2019 07:07:30",
      "content": "<p>Thanks a lot! From your findings, I update 7 whales that my model does not correctly predicted.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 474539,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "02/19/2019 14:22:23",
      "content": "<p>instead of predicting the text, for those whale id with text, i augment their images with text. text is treated as addition visual appearance features. I also create new fake image = empty + text</p>\n\n<p>an example of image that is difficult to match with whale image, bur easy with text</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/474539/11356/mtach_by_text.png\" alt=\"enter image description here\"></p>",
      "votes": null,
      "replies": [
        {
          "id": 474764,
          "author_name": "interneuron",
          "author_url": "",
          "post_date": "02/19/2019 19:15:34",
          "content": "<p>That’s really cool. I thought about doing something like that, possibly by chopping up the text-detection network using it in the whale network but I don’t really know how to do it. </p>\n\n<p>As for the ocr, I know it’s not the sort of thing the sponsors want as it’s unlikely users of their site will upload new images they have taken with added text matching something in the database, but matching the text to get a few more training examples seemed reasonable. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 475680,
          "author_name": "cooleel",
          "author_url": "",
          "post_date": "02/21/2019 04:09:20",
          "content": "<p>This is brilliant! Thanks for sharing. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "472894": "Following the excellent [kernel][1] by [Costas Voglis][2] I tried a little ocr on the images containing text using pytesseract. Since I cannot figure out how to get tesseract to work on kaggle I am posting the results here. \n\nAfter some preprocessing it does a decent job with the darker text though on the images with fainter text and strange font its not very good. Overlap between train and test is minimal, I found less than 50 images with the same #xxxx between them, but the data are the data and we must use what we have. There are probably more matches but I only looked at the ones with the #. In any case I strongly recommend verifying everything yourself.\nGood fortunes! \n\n\n  [1]: https://www.kaggle.com/voglinio/images-containing-text\n  [2]: https://www.kaggle.com/voglinio",
    "474291": "Thanks a lot! From your findings, I update 7 whales that my model does not correctly predicted.",
    "474539": "instead of predicting the text, for those whale id with text, i augment their images with text. text is treated as addition visual appearance features. I also create new fake image = empty + text\n\nan example of image that is difficult to match with whale image, bur easy with text\n\n   ![enter image description here][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/474539/11356/mtach_by_text.png",
    "474764": "That’s really cool. I thought about doing something like that, possibly by chopping up the text-detection network using it in the whale network but I don’t really know how to do it. \n\nAs for the ocr, I know it’s not the sort of thing the sponsors want as it’s unlikely users of their site will upload new images they have taken with added text matching something in the database, but matching the text to get a few more training examples seemed reasonable.",
    "475680": "This is brilliant! Thanks for sharing.",
    "1076117": "I write a script for a beginner who wants to Learn OCR Step by Step. Just go and watch and support me if you like:-\n\nhttps://www.kaggle.com/naim99/ocr-text-recognition-ocr-space-api-tesseract"
  },
  "source": "meta"
}