Hello Kagglers,
\nDid you notice that some images have text written on them, did that disturb your models?
", "messages": [ { "id": 2128701, "postDate": "2023-02-04T00:18:57.403Z", "content": "I usually try to get rid of texts that have been burned into the radiological images. I have read that those can disturb CNN's, but on the other hand that might have been only an issue related to a small size of the dataset.
", "rawMarkdown": "I usually try to get rid of texts that have been burned into the radiological images. I have read that those can disturb CNN's, but on the other hand that might have been only an issue related to a small size of the dataset.", "votes": 1 }, { "id": 2128255, "postDate": "2023-02-03T15:14:46.447Z", "content": "Someone could use GradCam (https://github.com/jacobgil/pytorch-grad-cam) to see if the text disturbs the model to recognize the right class.
", "rawMarkdown": "Someone could use GradCam (https://github.com/jacobgil/pytorch-grad-cam) to see if the text disturbs the model to recognize the right class.", "votes": 1 }, { "id": 2127525, "postDate": "2023-02-03T00:25:10.227Z", "content": "Hello Kagglers,
\nDid you notice that some images have text written on them, did that disturb your models?
", "rawMarkdown": "Hello Kagglers,\n\nDid you notice that some images have text written on them, did that disturb your models?", "votes": 1 }, { "id": 2128815, "postDate": "2023-02-04T05:30:18.653Z", "content": "It looks like those watermarks are at maximum pixel value, then you could remove it with temp=np.where(pixel<max(),pixel,0). As I used this temporal image to place landmarks to crop, it doesn't hurt the original maximum pixels.
", "rawMarkdown": "It looks like those watermarks are at maximum pixel value, then you could remove it with temp=np.where(pixel@vipin20 Perhaps by looking for the largest connected component?
", "rawMarkdown": "@vipin20 Perhaps by looking for the largest connected component?", "votes": 1 }, { "id": 2128553, "postDate": "2023-02-03T19:52:25.097Z", "content": "I would use a deep learning SOTA model for inpainting from Hugging face maybe. Check this one (https://github.com/Sanster/lama-cleaner). I believe it will give you better results than opencv inpaint.
", "rawMarkdown": "I would use a deep learning SOTA model for inpainting from Hugging face maybe. Check this one (https://github.com/Sanster/lama-cleaner). I believe it will give you better results than opencv inpaint. ", "votes": 1, "replies": [ { "id": 2129076, "postDate": "2023-02-04T09:49:18.490Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 2129077, "postDate": "2023-02-04T09:49:38.283Z", "content": "@davidbroberts wrote a notebook to remove those text markers
", "rawMarkdown": "@davidbroberts wrote [a notebook](https://www.kaggle.com/code/davidbroberts/mammography-remove-letter-markers) to remove those text markers\n" } ] }, { "id": 2127857, "postDate": "2023-02-03T08:55:40.667Z", "content": "Those texts explain which angles were used to accomplish the mammography. like the CC/MLO view
", "rawMarkdown": "Those texts explain which angles were used to accomplish the mammography. like the CC/MLO view", "replies": [ { "id": 2127861, "postDate": "2023-02-03T09:19:39.170Z", "content": "Yes thank you, but did you find that it somehow affects the models?
", "rawMarkdown": "Yes thank you, but did you find that it somehow affects the models? ", "replies": [ { "id": 2127907, "postDate": "2023-02-03T10:26:16.470Z", "content": "I think by cropping ROI you able to exclude this text signs from training. See eg https://www.kaggle.com/code/salmanahmedtamu/faster-dicom-loading-and-cropping
", "rawMarkdown": "I think by cropping ROI you able to exclude this text signs from training. See eg https://www.kaggle.com/code/salmanahmedtamu/faster-dicom-loading-and-cropping", "votes": 1 } ] } ] } ], "comments": [ { "id": 2128701, "author_name": "Antti Isosalo", "author_url": "", "post_date": "2023-02-04T00:18:57.403000", "content": "I usually try to get rid of texts that have been burned into the radiological images. I have read that those can disturb CNN's, but on the other hand that might have been only an issue related to a small size of the dataset.
", "votes": 1, "replies": [] }, { "id": 2128255, "author_name": "Ali Abdin", "author_url": "", "post_date": "2023-02-03T15:14:46.447000", "content": "Someone could use GradCam (https://github.com/jacobgil/pytorch-grad-cam) to see if the text disturbs the model to recognize the right class.
", "votes": 1, "replies": [] }, { "id": 2128815, "author_name": "Alfredo Maussa", "author_url": "", "post_date": "2023-02-04T05:30:18.653000", "content": "It looks like those watermarks are at maximum pixel value, then you could remove it with temp=np.where(pixel<max(),pixel,0). As I used this temporal image to place landmarks to crop, it doesn't hurt the original maximum pixels.
", "votes": 0, "replies": [] }, { "id": 2128057, "author_name": "Vipin Kumar", "author_url": "", "post_date": "2023-02-03T12:46:10.703000", "content": "Hi, @jarvisai7 any idea how can i remove text(watermark) from image in this dataset. I traid cv2.inpaint to remove watermark but after removing watermark, watermark area got blur.
\nhelp me to remove text!
\nThanks
@vipin20 Perhaps by looking for the largest connected component?
", "votes": 1, "replies": [] }, { "id": 2128553, "author_name": "Yacine Bouaouni", "author_url": "", "post_date": "2023-02-03T19:52:25.097000", "content": "I would use a deep learning SOTA model for inpainting from Hugging face maybe. Check this one (https://github.com/Sanster/lama-cleaner). I believe it will give you better results than opencv inpaint.
", "votes": 1, "replies": [ { "id": 2129076, "author_name": "", "author_url": "", "post_date": "2023-02-04T09:49:18.490000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 2129077, "author_name": "FabienDaniel", "author_url": "", "post_date": "2023-02-04T09:49:38.283000", "content": "@davidbroberts wrote a notebook to remove those text markers
", "votes": 0, "replies": [] } ] }, { "id": 2127857, "author_name": "Rahel Sarif", "author_url": "", "post_date": "2023-02-03T08:55:40.667000", "content": "Those texts explain which angles were used to accomplish the mammography. like the CC/MLO view
", "votes": 0, "replies": [ { "id": 2127861, "author_name": "Yacine Bouaouni", "author_url": "", "post_date": "2023-02-03T09:19:39.170000", "content": "Yes thank you, but did you find that it somehow affects the models?
", "votes": 0, "replies": [ { "id": 2127907, "author_name": "A.P.", "author_url": "", "post_date": "2023-02-03T10:26:16.470000", "content": "I think by cropping ROI you able to exclude this text signs from training. See eg https://www.kaggle.com/code/salmanahmedtamu/faster-dicom-loading-and-cropping
", "votes": 1, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2128701": "I usually try to get rid of texts that have been burned into the radiological images. I have read that those can disturb CNN's, but on the other hand that might have been only an issue related to a small size of the dataset.", "2128255": "Someone could use GradCam (https://github.com/jacobgil/pytorch-grad-cam) to see if the text disturbs the model to recognize the right class.", "2127525": "Hello Kagglers,\n\nDid you notice that some images have text written on them, did that disturb your models?", "2128815": "It looks like those watermarks are at maximum pixel value, then you could remove it with temp=np.where(pixel