{
  "id": 15045,
  "title": "Image Normalization",
  "url": "/competitions/diabetic-retinopathy-detection/discussion/15045",
  "author_name": "",
  "post_date": "2015-07-04T17:35:44.483Z",
  "votes": null,
  "comment_count": 4,
  "views": 1723,
  "content": "<p>Dear Contestants,</p>\n<p>I have managed to implement several vessel extraction techniques, macula, OTN, and lesions in the images, however, for many images they won't work. The reason is that the images are not normalized, they are of low contrast, low lighting and similar things.</p>\n<p>I wonder how can we manage to normalize the images? or remove the noise in the images?</p>\n<p>Thanks.</p>",
  "messages": [
    {
      "id": "83429",
      "postDate": "07/04/2015 17:35:44",
      "content": "<p>Dear Contestants,</p>\n<p>I have managed to implement several vessel extraction techniques, macula, OTN, and lesions in the images, however, for many images they won't work. The reason is that the images are not normalized, they are of low contrast, low lighting and similar things.</p>\n<p>I wonder how can we manage to normalize the images? or remove the noise in the images?</p>\n<p>Thanks.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "83448",
      "postDate": "07/05/2015 00:01:05",
      "content": "<p>Check this link&nbsp;http://www.goapi.upc.edu/usr/andre/sv-restoration/index.html&nbsp; the key is &quot;deconvolution&quot;&nbsp;<a href=\"http://biomedicaloptics.spiedigitallibrary.org/article.aspx?articleid=1167113\" target=\"_blank\">http://biomedicaloptics.spiedigitallibrary.org/article.aspx?articleid=1167113</a></p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "83661",
      "postDate": "07/07/2015 13:48:46",
      "content": "<p>Although those deconvolution concepts are applying to the same retina, taken at different time periods - vs random different images of different eyes, if that sort of thing bothers you.\n(i.e. it follows that taking different data points of the same thing and trying to find less noise would work, but does it follow that taking different data points of different things will result in less noise, or just &quot;different noise&quot;)</p>\n\n<p>To me, that technique reminds me of signal boosting in audio - where you play two noisy signals next to each other (example two cop cars next to each other talking, and the same noisy signal comes over the radio - if you are standing there, you will get a cleaner and louder signal from both combined than you would either one individually) the signals double, and the noise either cancel out, or lose out to the boosted signal. So the equivalent veins would boost their signal on the same eye, and then new noise elements like lesions that weren't there before would stand out.\nBut if you took my eye, and then took your eye - are our veins likely to line up so well so as to allow for that lowering of noise?\nGiven there is a whole retinal recognition thing out there, it leads me to think that retinas from different eyes are different enough that you cannot use two random retinas to help get that effect.</p>\n\n<p>Keeping in mind that maybe I totally misunderstood what they were talking about and I am now just babbling and waving my arms around like a crazy person.</p>",
      "rawMarkdown": "Although those deconvolution concepts are applying to the same retina, taken at different time periods - vs random different images of different eyes, if that sort of thing bothers you.\r\n(i.e. it follows that taking different data points of the same thing and trying to find less noise would work, but does it follow that taking different data points of different things will result in less noise, or just \"different noise\")\r\n\r\nTo me, that technique reminds me of signal boosting in audio - where you play two noisy signals next to each other (example two cop cars next to each other talking, and the same noisy signal comes over the radio - if you are standing there, you will get a cleaner and louder signal from both combined than you would either one individually) the signals double, and the noise either cancel out, or lose out to the boosted signal. So the equivalent veins would boost their signal on the same eye, and then new noise elements like lesions that weren't there before would stand out.\r\nBut if you took my eye, and then took your eye - are our veins likely to line up so well so as to allow for that lowering of noise?\r\nGiven there is a whole retinal recognition thing out there, it leads me to think that retinas from different eyes are different enough that you cannot use two random retinas to help get that effect.\r\n\r\nKeeping in mind that maybe I totally misunderstood what they were talking about and I am now just babbling and waving my arms around like a crazy person.",
      "votes": null
    },
    {
      "id": "86636",
      "postDate": "07/23/2015 14:34:06",
      "content": "<p>You can use adaptive histogram equalization to improve the contrast and reduce the noise. </p>\n\n<p>To extract the blood vessels use the matched filter method mentioned elsewhere in these posts.</p>\n\n<p>Jals123</p>",
      "rawMarkdown": "You can use adaptive histogram equalization to improve the contrast and reduce the noise. \r\n\r\nTo extract the blood vessels use the matched filter method mentioned elsewhere in these posts.\r\n\r\nJals123",
      "votes": null
    },
    {
      "id": "87062",
      "postDate": "07/26/2015 18:16:44",
      "content": "<p>Did not help me in my ConvNet, but works like a charm on its own:</p>\n\n<p><a href=\"http://www.graphicsmagick.org/GraphicsMagick.html#details-normalize\">http://www.graphicsmagick.org/GraphicsMagick.html#details-normalize</a></p>",
      "rawMarkdown": "Did not help me in my ConvNet, but works like a charm on its own:\r\n\r\nhttp://www.graphicsmagick.org/GraphicsMagick.html#details-normalize",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 83448,
      "author_name": "eduinhserna",
      "author_url": "",
      "post_date": "07/05/2015 00:01:05",
      "content": "<p>Check this link&nbsp;http://www.goapi.upc.edu/usr/andre/sv-restoration/index.html&nbsp; the key is &quot;deconvolution&quot;&nbsp;<a href=\"http://biomedicaloptics.spiedigitallibrary.org/article.aspx?articleid=1167113\" target=\"_blank\">http://biomedicaloptics.spiedigitallibrary.org/article.aspx?articleid=1167113</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 83661,
      "author_name": "omgponies",
      "author_url": "",
      "post_date": "07/07/2015 13:48:46",
      "content": "<p>Although those deconvolution concepts are applying to the same retina, taken at different time periods - vs random different images of different eyes, if that sort of thing bothers you.\n(i.e. it follows that taking different data points of the same thing and trying to find less noise would work, but does it follow that taking different data points of different things will result in less noise, or just &quot;different noise&quot;)</p>\n\n<p>To me, that technique reminds me of signal boosting in audio - where you play two noisy signals next to each other (example two cop cars next to each other talking, and the same noisy signal comes over the radio - if you are standing there, you will get a cleaner and louder signal from both combined than you would either one individually) the signals double, and the noise either cancel out, or lose out to the boosted signal. So the equivalent veins would boost their signal on the same eye, and then new noise elements like lesions that weren't there before would stand out.\nBut if you took my eye, and then took your eye - are our veins likely to line up so well so as to allow for that lowering of noise?\nGiven there is a whole retinal recognition thing out there, it leads me to think that retinas from different eyes are different enough that you cannot use two random retinas to help get that effect.</p>\n\n<p>Keeping in mind that maybe I totally misunderstood what they were talking about and I am now just babbling and waving my arms around like a crazy person.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 86636,
      "author_name": "jals123",
      "author_url": "",
      "post_date": "07/23/2015 14:34:06",
      "content": "<p>You can use adaptive histogram equalization to improve the contrast and reduce the noise. </p>\n\n<p>To extract the blood vessels use the matched filter method mentioned elsewhere in these posts.</p>\n\n<p>Jals123</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 87062,
      "author_name": "ilyakava",
      "author_url": "",
      "post_date": "07/26/2015 18:16:44",
      "content": "<p>Did not help me in my ConvNet, but works like a charm on its own:</p>\n\n<p><a href=\"http://www.graphicsmagick.org/GraphicsMagick.html#details-normalize\">http://www.graphicsmagick.org/GraphicsMagick.html#details-normalize</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "83429": "",
    "83448": "",
    "83661": "Although those deconvolution concepts are applying to the same retina, taken at different time periods - vs random different images of different eyes, if that sort of thing bothers you.\r\n(i.e. it follows that taking different data points of the same thing and trying to find less noise would work, but does it follow that taking different data points of different things will result in less noise, or just \"different noise\")\r\n\r\nTo me, that technique reminds me of signal boosting in audio - where you play two noisy signals next to each other (example two cop cars next to each other talking, and the same noisy signal comes over the radio - if you are standing there, you will get a cleaner and louder signal from both combined than you would either one individually) the signals double, and the noise either cancel out, or lose out to the boosted signal. So the equivalent veins would boost their signal on the same eye, and then new noise elements like lesions that weren't there before would stand out.\r\nBut if you took my eye, and then took your eye - are our veins likely to line up so well so as to allow for that lowering of noise?\r\nGiven there is a whole retinal recognition thing out there, it leads me to think that retinas from different eyes are different enough that you cannot use two random retinas to help get that effect.\r\n\r\nKeeping in mind that maybe I totally misunderstood what they were talking about and I am now just babbling and waving my arms around like a crazy person.",
    "86636": "You can use adaptive histogram equalization to improve the contrast and reduce the noise. \r\n\r\nTo extract the blood vessels use the matched filter method mentioned elsewhere in these posts.\r\n\r\nJals123",
    "87062": "Did not help me in my ConvNet, but works like a charm on its own:\r\n\r\nhttp://www.graphicsmagick.org/GraphicsMagick.html#details-normalize"
  },
  "source": "meta"
}