![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F745525%2F46eb2e21f3236c3d7c52fb9342d4d917%2F2020-01-29%2023.43.47.png?generation=1580309063628317&alt=media\")
This kernel shows some image augmentation using random morphological transformations. It might help to make the model to be more robust. Or standard augmentation might be enough... Anyway try it!
\n\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F745525%2F1eb01b19ef93ebc9fde1ae44751e8a4f%2F2020-01-29%2023.43.56.png?generation=1580309066652181&alt=media\")
This kernel shows some image augmentation using random morphological transformations. It might help to make the model to be more robust. Or standard augmentation might be enough... Anyway try it!
\n\n
Really cool idea, excellent variation for augmenting the data!
", "rawMarkdown": "Really cool idea, excellent variation for augmenting the data!", "votes": null }, { "id": "734220", "postDate": "02/01/2020 05:29:33", "content": "Thank you for sharing an interesting idea. \nI wrote a basic albumetations' transform using cv2.erode and cv2.dilate from the kernel.
```\nclass RandomMorph(ImageOnlyTransform):
\n\ndef __init__(self, _min=2, _max=6, element_shape=cv2.MORPH_ELLIPSE, always_apply=False, p=0.5):\n super().__init__(always_apply, p)\n self._min = _min\n self._max = _max\n self.element_shape = element_shape\n\ndef apply(self, image, **params):\n arr = np.random.randint(self._min, self._max, 2)\n kernel = cv2.getStructuringElement(self.element_shape, tuple(arr))\n\n if random.random() > 0.5:\n # make it thinner\n image = cv2.erode(image, kernel, iterations=1)\n else:\n # make it thicker\n image = cv2.dilate(image, kernel, iterations=1)\n\n return image\n```
", "rawMarkdown": "Thank you for sharing an interesting idea. \nI wrote a basic albumetations' transform using `cv2.erode` and `cv2.dilate` from [the kernel](https://www.kaggle.com/ren4yu/bengali-morphological-ops-as-image-augmentation).\n\n```\nclass RandomMorph(ImageOnlyTransform):\n\n def __init__(self, _min=2, _max=6, element_shape=cv2.MORPH_ELLIPSE, always_apply=False, p=0.5):\n super().__init__(always_apply, p)\n self._min = _min\n self._max = _max\n self.element_shape = element_shape\n\n def apply(self, image, **params):\n arr = np.random.randint(self._min, self._max, 2)\n kernel = cv2.getStructuringElement(self.element_shape, tuple(arr))\n\n if random.random() > 0.5:\n # make it thinner\n image = cv2.erode(image, kernel, iterations=1)\n else:\n # make it thicker\n image = cv2.dilate(image, kernel, iterations=1)\n\n return image\n```", "votes": null }, { "id": "734927", "postDate": "02/02/2020 08:32:11", "content": "Thank you!
", "rawMarkdown": "Thank you!", "votes": null }, { "id": "751492", "postDate": "02/20/2020 08:28:31", "content": "Nice
", "rawMarkdown": "Nice", "votes": null }, { "id": "752678", "postDate": "02/21/2020 09:46:50", "content": "good
", "rawMarkdown": "good", "votes": null } ], "comments": [ { "id": 733114, "author_name": "ninapberry", "author_url": "", "post_date": "01/30/2020 17:18:40", "content": "Really cool idea, excellent variation for augmenting the data!
", "votes": null, "replies": [] }, { "id": 734220, "author_name": "appian", "author_url": "", "post_date": "02/01/2020 05:29:33", "content": "Thank you for sharing an interesting idea. \nI wrote a basic albumetations' transform using cv2.erode and cv2.dilate from the kernel.
```\nclass RandomMorph(ImageOnlyTransform):
\n\ndef __init__(self, _min=2, _max=6, element_shape=cv2.MORPH_ELLIPSE, always_apply=False, p=0.5):\n super().__init__(always_apply, p)\n self._min = _min\n self._max = _max\n self.element_shape = element_shape\n\ndef apply(self, image, **params):\n arr = np.random.randint(self._min, self._max, 2)\n kernel = cv2.getStructuringElement(self.element_shape, tuple(arr))\n\n if random.random() > 0.5:\n # make it thinner\n image = cv2.erode(image, kernel, iterations=1)\n else:\n # make it thicker\n image = cv2.dilate(image, kernel, iterations=1)\n\n return image\n```
", "votes": null, "replies": [ { "id": 734927, "author_name": "moximo13", "author_url": "", "post_date": "02/02/2020 08:32:11", "content": "Thank you!
", "votes": null, "replies": [] } ] }, { "id": 751492, "author_name": "", "author_url": "", "post_date": "02/20/2020 08:28:31", "content": "Nice
", "votes": null, "replies": [] }, { "id": 752678, "author_name": "", "author_url": "", "post_date": "02/21/2020 09:46:50", "content": "good
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "732182": "[This kernel](https://www.kaggle.com/ren4yu/bengali-morphological-ops-as-image-augmentation) shows some image augmentation using random morphological transformations. It might help to make the model to be more robust. Or standard augmentation might be enough... Anyway try it!\n\n\n\n", "733114": "Really cool idea, excellent variation for augmenting the data!", "734220": "Thank you for sharing an interesting idea. \nI wrote a basic albumetations' transform using `cv2.erode` and `cv2.dilate` from [the kernel](https://www.kaggle.com/ren4yu/bengali-morphological-ops-as-image-augmentation).\n\n```\nclass RandomMorph(ImageOnlyTransform):\n\n def __init__(self, _min=2, _max=6, element_shape=cv2.MORPH_ELLIPSE, always_apply=False, p=0.5):\n super().__init__(always_apply, p)\n self._min = _min\n self._max = _max\n self.element_shape = element_shape\n\n def apply(self, image, **params):\n arr = np.random.randint(self._min, self._max, 2)\n kernel = cv2.getStructuringElement(self.element_shape, tuple(arr))\n\n if random.random() > 0.5:\n # make it thinner\n image = cv2.erode(image, kernel, iterations=1)\n else:\n # make it thicker\n image = cv2.dilate(image, kernel, iterations=1)\n\n return image\n```", "734927": "Thank you!", "751492": "Nice", "752678": "good" }, "source": "meta" }