{
  "id": 149238,
  "title": "How does Data Augmentation and transformations affects on Steganalysis?",
  "url": "/competitions/alaska2-image-steganalysis/discussion/149238",
  "author_name": "",
  "post_date": "2020-05-07T12:37:18.568959900Z",
  "votes": 4,
  "comment_count": 2,
  "views": 0,
  "content": "<p>My first thought when I started using NN was if I should resize the input images or stick to the 512x512 format and if I should apply data augmentation to it. Has anyone been looking into this?</p>\n\n<p>I've seen lots of Notebook in this challenge where images are resized. Of course it makes the training faster while using less parameters. However, my thoughts on why not resizing the input images is because the interpolation of the pixels when resizing affects greatly the high frequencies (high DCT coefficients), thus losing part or changing most part payload of the encrypted data inside the image. My fear is that resizing might make training worse.</p>\n\n<p>Same as data augmentation. All data augmentation might corrupt the payload but there are some techniques that do not corrupt entirely like flips. As we have 3 time more encrypted data should we apply data augmentation?</p>\n\n<p>What do you think?</p>",
  "messages": [
    {
      "id": "837004",
      "postDate": "05/07/2020 12:37:18",
      "content": "<p>My first thought when I started using NN was if I should resize the input images or stick to the 512x512 format and if I should apply data augmentation to it. Has anyone been looking into this?</p>\n\n<p>I've seen lots of Notebook in this challenge where images are resized. Of course it makes the training faster while using less parameters. However, my thoughts on why not resizing the input images is because the interpolation of the pixels when resizing affects greatly the high frequencies (high DCT coefficients), thus losing part or changing most part payload of the encrypted data inside the image. My fear is that resizing might make training worse.</p>\n\n<p>Same as data augmentation. All data augmentation might corrupt the payload but there are some techniques that do not corrupt entirely like flips. As we have 3 time more encrypted data should we apply data augmentation?</p>\n\n<p>What do you think?</p>",
      "rawMarkdown": "My first thought when I started using NN was if I should resize the input images or stick to the 512x512 format and if I should apply data augmentation to it. Has anyone been looking into this?\n\nI've seen lots of Notebook in this challenge where images are resized. Of course it makes the training faster while using less parameters. However, my thoughts on why not resizing the input images is because the interpolation of the pixels when resizing affects greatly the high frequencies (high DCT coefficients), thus losing part or changing most part payload of the encrypted data inside the image. My fear is that resizing might make training worse.\n\nSame as data augmentation. All data augmentation might corrupt the payload but there are some techniques that do not corrupt entirely like flips. As we have 3 time more encrypted data should we apply data augmentation?\n\nWhat do you think?",
      "votes": null
    },
    {
      "id": "838242",
      "postDate": "05/08/2020 12:08:25",
      "content": "<p>I'd say your assumptions are correct. Careless image resizing / augmentation will damage those hi-frequency artifacts. Therefore one should end up with augmentation techniques that prevent over-fitting and don't change data distribution at the same time.</p>",
      "rawMarkdown": "I'd say your assumptions are correct. Careless image resizing / augmentation will damage those hi-frequency artifacts. Therefore one should end up with augmentation techniques that prevent over-fitting and don't change data distribution at the same time.",
      "votes": null
    },
    {
      "id": "842373",
      "postDate": "05/11/2020 11:09:48",
      "content": "<p>Hello there,</p>\n\n<p>Very good answer.\nResizing will completely obliterate all changes made in the DCT coefficient (and/or in pixels).\nBest way for data aumgnetation ([see this bright paper on this topic](other trick that you guys know much better than me such as )) \n- Crop \n- Rotate (without interpolation, so only by 90, 180, 270 degre)\n- Flip\n....</p>",
      "rawMarkdown": "Hello there,\n\nVery good answer.\nResizing will completely obliterate all changes made in the DCT coefficient (and/or in pixels).\nBest way for data aumgnetation ([see this bright paper on this topic](other trick that you guys know much better than me such as )) \n- Crop \n- Rotate (without interpolation, so only by 90, 180, 270 degre)\n- Flip\n....",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 838242,
      "author_name": "bloodaxe",
      "author_url": "",
      "post_date": "05/08/2020 12:08:25",
      "content": "<p>I'd say your assumptions are correct. Careless image resizing / augmentation will damage those hi-frequency artifacts. Therefore one should end up with augmentation techniques that prevent over-fitting and don't change data distribution at the same time.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 842373,
      "author_name": "remicogranne",
      "author_url": "",
      "post_date": "05/11/2020 11:09:48",
      "content": "<p>Hello there,</p>\n\n<p>Very good answer.\nResizing will completely obliterate all changes made in the DCT coefficient (and/or in pixels).\nBest way for data aumgnetation ([see this bright paper on this topic](other trick that you guys know much better than me such as )) \n- Crop \n- Rotate (without interpolation, so only by 90, 180, 270 degre)\n- Flip\n....</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "837004": "My first thought when I started using NN was if I should resize the input images or stick to the 512x512 format and if I should apply data augmentation to it. Has anyone been looking into this?\n\nI've seen lots of Notebook in this challenge where images are resized. Of course it makes the training faster while using less parameters. However, my thoughts on why not resizing the input images is because the interpolation of the pixels when resizing affects greatly the high frequencies (high DCT coefficients), thus losing part or changing most part payload of the encrypted data inside the image. My fear is that resizing might make training worse.\n\nSame as data augmentation. All data augmentation might corrupt the payload but there are some techniques that do not corrupt entirely like flips. As we have 3 time more encrypted data should we apply data augmentation?\n\nWhat do you think?",
    "838242": "I'd say your assumptions are correct. Careless image resizing / augmentation will damage those hi-frequency artifacts. Therefore one should end up with augmentation techniques that prevent over-fitting and don't change data distribution at the same time.",
    "842373": "Hello there,\n\nVery good answer.\nResizing will completely obliterate all changes made in the DCT coefficient (and/or in pixels).\nBest way for data aumgnetation ([see this bright paper on this topic](other trick that you guys know much better than me such as )) \n- Crop \n- Rotate (without interpolation, so only by 90, 180, 270 degre)\n- Flip\n...."
  },
  "source": "meta"
}