{
  "id": 158861,
  "title": "What image augmentation methods worked best?",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/158861",
  "author_name": "",
  "post_date": "2020-06-15T16:15:08.562758900Z",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Would love to see a thread of which image augmentation methods worked best for your models.</p>",
  "messages": [
    {
      "id": "887349",
      "postDate": "06/15/2020 16:15:08",
      "content": "<p>Would love to see a thread of which image augmentation methods worked best for your models.</p>",
      "rawMarkdown": "Would love to see a thread of which image augmentation methods worked best for your models.",
      "votes": null
    },
    {
      "id": "887864",
      "postDate": "06/16/2020 00:22:43",
      "content": "<p>I have used coarse dropout + some flips + changing brightness etc. Basically the most popular stuff from albumentations :). Right now, I am implementing some augmentations from that <a href=\"https://l.facebook.com/l.php?u=https%3A%2F%2Farxiv.org%2Fpdf%2F1809.02568.pdf%3Ffbclid%3DIwAR0FaE5oK1criWNQKc7cKFIq3xAWHOaPZkUIuj6QAcOIsDe0XD1W9yBfUzc&amp;h=AT1iiEjPlcJWE-7o2Nom1GK4QFWJcR2kdfzJTmPlUR4nJr-PGhkx6huSnOIA9CLmCyTjrw1TUNF5MLEwsieJD6i44Gu2hPUt38ggCTDXNVom9vpybMYitBPV9rOJ57Ei5GUl\">paper</a>.</p>",
      "rawMarkdown": "I have used coarse dropout + some flips + changing brightness etc. Basically the most popular stuff from albumentations :). Right now, I am implementing some augmentations from that [paper](https://l.facebook.com/l.php?u=https%3A%2F%2Farxiv.org%2Fpdf%2F1809.02568.pdf%3Ffbclid%3DIwAR0FaE5oK1criWNQKc7cKFIq3xAWHOaPZkUIuj6QAcOIsDe0XD1W9yBfUzc&amp;h=AT1iiEjPlcJWE-7o2Nom1GK4QFWJcR2kdfzJTmPlUR4nJr-PGhkx6huSnOIA9CLmCyTjrw1TUNF5MLEwsieJD6i44Gu2hPUt38ggCTDXNVom9vpybMYitBPV9rOJ57Ei5GUl).",
      "votes": null
    },
    {
      "id": "887998",
      "postDate": "06/16/2020 03:56:19",
      "content": "<p>That's great. Thank you for sharing the paper.</p>",
      "rawMarkdown": "That's great. Thank you for sharing the paper.",
      "votes": null
    },
    {
      "id": "891559",
      "postDate": "06/18/2020 08:57:41",
      "content": "<p>Image denoising is of paramount importance. I suggest to go with either Morphological processing opening operation(where we erode the image first, followed by dilation) or  Non-Local Means Denoising. It takes more time compared to blurring techniques, but the result are very satisfying. </p>\n\n<p>I have done it using the latter approach here </p>\n\n<p><em><strong><a href=\"https://www.kaggle.com/fireheart7/melanoma-eda-cum-preprocessing\">https://www.kaggle.com/fireheart7/melanoma-eda-cum-preprocessing</a></strong></em></p>\n\n<p>Moreover, I would suggest to balance out the intensity by local histogram equalization. Please note that Global histogram equalization may fail here as the area of interest is actually small in many images. So, it might be just possible that those intensities gets further suppressed. </p>\n\n<p>Both these methods along with Kmeans image segmentation are performed in the above kernel. Please consider upvoting if you find these insights useful.</p>\n\n<p>Thank you and all the best! </p>",
      "rawMarkdown": "Image denoising is of paramount importance. I suggest to go with either Morphological processing opening operation(where we erode the image first, followed by dilation) or  Non-Local Means Denoising. It takes more time compared to blurring techniques, but the result are very satisfying. \n\nI have done it using the latter approach here \n\n***https://www.kaggle.com/fireheart7/melanoma-eda-cum-preprocessing***\n\nMoreover, I would suggest to balance out the intensity by local histogram equalization. Please note that Global histogram equalization may fail here as the area of interest is actually small in many images. So, it might be just possible that those intensities gets further suppressed. \n\nBoth these methods along with Kmeans image segmentation are performed in the above kernel. Please consider upvoting if you find these insights useful.\n\nThank you and all the best!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 887864,
      "author_name": "janmalin",
      "author_url": "",
      "post_date": "06/16/2020 00:22:43",
      "content": "<p>I have used coarse dropout + some flips + changing brightness etc. Basically the most popular stuff from albumentations :). Right now, I am implementing some augmentations from that <a href=\"https://l.facebook.com/l.php?u=https%3A%2F%2Farxiv.org%2Fpdf%2F1809.02568.pdf%3Ffbclid%3DIwAR0FaE5oK1criWNQKc7cKFIq3xAWHOaPZkUIuj6QAcOIsDe0XD1W9yBfUzc&amp;h=AT1iiEjPlcJWE-7o2Nom1GK4QFWJcR2kdfzJTmPlUR4nJr-PGhkx6huSnOIA9CLmCyTjrw1TUNF5MLEwsieJD6i44Gu2hPUt38ggCTDXNVom9vpybMYitBPV9rOJ57Ei5GUl\">paper</a>.</p>",
      "votes": null,
      "replies": [
        {
          "id": 887998,
          "author_name": "pragyanbo",
          "author_url": "",
          "post_date": "06/16/2020 03:56:19",
          "content": "<p>That's great. Thank you for sharing the paper.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 891559,
      "author_name": "fireheart7",
      "author_url": "",
      "post_date": "06/18/2020 08:57:41",
      "content": "<p>Image denoising is of paramount importance. I suggest to go with either Morphological processing opening operation(where we erode the image first, followed by dilation) or  Non-Local Means Denoising. It takes more time compared to blurring techniques, but the result are very satisfying. </p>\n\n<p>I have done it using the latter approach here </p>\n\n<p><em><strong><a href=\"https://www.kaggle.com/fireheart7/melanoma-eda-cum-preprocessing\">https://www.kaggle.com/fireheart7/melanoma-eda-cum-preprocessing</a></strong></em></p>\n\n<p>Moreover, I would suggest to balance out the intensity by local histogram equalization. Please note that Global histogram equalization may fail here as the area of interest is actually small in many images. So, it might be just possible that those intensities gets further suppressed. </p>\n\n<p>Both these methods along with Kmeans image segmentation are performed in the above kernel. Please consider upvoting if you find these insights useful.</p>\n\n<p>Thank you and all the best! </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "887349": "Would love to see a thread of which image augmentation methods worked best for your models.",
    "887864": "I have used coarse dropout + some flips + changing brightness etc. Basically the most popular stuff from albumentations :). Right now, I am implementing some augmentations from that [paper](https://l.facebook.com/l.php?u=https%3A%2F%2Farxiv.org%2Fpdf%2F1809.02568.pdf%3Ffbclid%3DIwAR0FaE5oK1criWNQKc7cKFIq3xAWHOaPZkUIuj6QAcOIsDe0XD1W9yBfUzc&amp;h=AT1iiEjPlcJWE-7o2Nom1GK4QFWJcR2kdfzJTmPlUR4nJr-PGhkx6huSnOIA9CLmCyTjrw1TUNF5MLEwsieJD6i44Gu2hPUt38ggCTDXNVom9vpybMYitBPV9rOJ57Ei5GUl).",
    "887998": "That's great. Thank you for sharing the paper.",
    "891559": "Image denoising is of paramount importance. I suggest to go with either Morphological processing opening operation(where we erode the image first, followed by dilation) or  Non-Local Means Denoising. It takes more time compared to blurring techniques, but the result are very satisfying. \n\nI have done it using the latter approach here \n\n***https://www.kaggle.com/fireheart7/melanoma-eda-cum-preprocessing***\n\nMoreover, I would suggest to balance out the intensity by local histogram equalization. Please note that Global histogram equalization may fail here as the area of interest is actually small in many images. So, it might be just possible that those intensities gets further suppressed. \n\nBoth these methods along with Kmeans image segmentation are performed in the above kernel. Please consider upvoting if you find these insights useful.\n\nThank you and all the best!"
  },
  "source": "meta"
}