binarization first. Our group is working on this topic. For better understanding document binarization, please see some samples bellow: \nRaw image \n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1938879%2F3ffc9572d64131d47b7ecdc4e99bc267%2F100241706_00015_2.jpg?generation=1568273728886431&alt=media\")
Hi all, \nI wish I have time to take part in this competition because I want to check if our research topic can help in this kind of task. Regarding to document processing, especially hand-written document, we often do binarization first. Our group is working on this topic. For better understanding document binarization, please see some samples bellow: \nRaw image \n
Binarized image \n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1938879%2F89c7704ff1744c6d8591fe1eda116d99%2F100241706_00015_2.jpg?generation=1568273756500375&alt=media\")
Raw image \ntest_0ad567f7.jpg\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1938879%2Fb814b7410d7f799254eb7bf1cc95b1fe%2Ftest_0ad567f7.jpg?generation=1568273925103956&alt=media\")
Binarized image \ntest_0ad567f7.jpg \n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1938879%2F07f028eec3f8c9fbdabb3e4945e2a8a0%2Ftest_0ad567f7.jpg?generation=1568273972570781&alt=media\")
Those images are not that bad right?
\n\nThey are processed by our pretrained models which using datasets: DIBCO, H-DIBCO, etc. \nEvent those datasets above are different from Kuzushiji dataset, however, I am surprised that the model can perform well.
\n\nI upload binarized dataset in here: \nhttps://www.kaggle.com/backaggle/kuzushijibin
Please let me know if the binarized dataset can help you. \nHappy Kaggling
", "messages": [ { "id": "624670", "postDate": "09/12/2019 09:06:20", "content": "Hi all, \nI wish I have time to take part in this competition because I want to check if our research topic can help in this kind of task. Regarding to document processing, especially hand-written document, we often do binarization first. Our group is working on this topic. For better understanding document binarization, please see some samples bellow: \nRaw image \n
Binarized image \n
Raw image \ntest_0ad567f7.jpg\n
Binarized image \ntest_0ad567f7.jpg \n
Those images are not that bad right?
\n\nThey are processed by our pretrained models which using datasets: DIBCO, H-DIBCO, etc. \nEvent those datasets above are different from Kuzushiji dataset, however, I am surprised that the model can perform well.
\n\nI upload binarized dataset in here: \nhttps://www.kaggle.com/backaggle/kuzushijibin
Please let me know if the binarized dataset can help you. \nHappy Kaggling
", "rawMarkdown": "Hi all, \nI wish I have time to take part in this competition because I want to check if our research topic can help in this kind of task. Regarding to document processing, especially hand-written document, we often do `binarization` first. Our group is working on this topic. For better understanding document binarization, please see some samples bellow: \n**Raw image** \n\n\n\n**Binarized image** \n\n\n\n**Raw image** \ntest_0ad567f7.jpg\n\n\n\n\n**Binarized image** \ntest_0ad567f7.jpg \n\n\nThose images are not that bad right? \n\nThey are processed by our pretrained models which using datasets: DIBCO, H-DIBCO, etc. \nEvent those datasets above are different from Kuzushiji dataset, however, I am surprised that the model can perform well. \n\nI upload `binarized dataset` in here: \nhttps://www.kaggle.com/backaggle/kuzushijibin \n\nPlease let me know if the binarized dataset can help you. \nHappy Kaggling", "votes": null }, { "id": "624703", "postDate": "09/12/2019 09:47:42", "content": "Hi, thanks for bringing it up, I’m also curious if binarization will help here. Are binarization models which you used public? As I understand, we won’t be able to the data you posted for submissions if models aren’t public under some appropriate license.
", "rawMarkdown": "Hi, thanks for bringing it up, I’m also curious if binarization will help here. Are binarization models which you used public? As I understand, we won’t be able to the data you posted for submissions if models aren’t public under some appropriate license.", "votes": null }, { "id": "624708", "postDate": "09/12/2019 09:54:19", "content": "Our models have been not published yet. They are still under research. However, I think you can use this model: \nhttps://github.com/masyagin1998/robin
\n\nOur performance is not much different compared to this. I tried this model and the binarized outputs are quite similar. It should not affect the results.
", "rawMarkdown": "Our models have been not published yet. They are still under research. However, I think you can use this model: \nhttps://github.com/masyagin1998/robin \n\nOur performance is not much different compared to this. I tried this model and the binarized outputs are quite similar. It should not affect the results.", "votes": null }, { "id": "625039", "postDate": "09/12/2019 16:43:21", "content": "Hi @backaggle, your binarization looks great! Could you give us a clue about your approach (algorithm, strategy, ...)?
\n\nThank you.\nJesús
", "rawMarkdown": "Hi @backaggle, your binarization looks great! Could you give us a clue about your approach (algorithm, strategy, ...)?\n\nThank you.\nJesús", "votes": null }, { "id": "625363", "postDate": "09/13/2019 01:48:00", "content": "Binarization can be treated as a segmentation problem where you try to classify whether a pixel belongs to the strokes or the background. We use typical segmentation models like: Unet, Deeplab, etc.
One of the problems in this field is the lack of data. The published datasets are totally a few hundred. in both hand-writing and printing type. It is easy to make a ground truth, but it is hard to make a training data because you should make the background looks real. To deal with this problem, we use Cycle-GAN to transfer the background style to the different ground truths. We were successful in this approach.
Interesting topic. Are you going to release anytime soon your algos?
", "rawMarkdown": "Interesting topic. Are you going to release anytime soon your algos?", "votes": null }, { "id": "626216", "postDate": "09/14/2019 03:08:56", "content": "Thank for your interest. \nAs soon as we have the result of DIBCO competition this year, we will release our paper.
", "rawMarkdown": "Thank for your interest. \nAs soon as we have the result of DIBCO competition this year, we will release our paper.", "votes": null }, { "id": "2466689", "postDate": "10/04/2023 03:42:02", "content": "Dear Sir, I need complete code of entitled \"Three-stage binarization of color document images based on discrete wavelet transform and generative adversarial networks.\" to complete my master project.
", "rawMarkdown": "Dear Sir, I need complete code of entitled \"Three-stage binarization of color document images based on discrete wavelet transform and generative adversarial networks.\" to complete my master project.", "votes": null } ], "comments": [ { "id": 624703, "author_name": "lopuhin", "author_url": "", "post_date": "09/12/2019 09:47:42", "content": "Hi, thanks for bringing it up, I’m also curious if binarization will help here. Are binarization models which you used public? As I understand, we won’t be able to the data you posted for submissions if models aren’t public under some appropriate license.
", "votes": null, "replies": [ { "id": 624708, "author_name": "backaggle", "author_url": "", "post_date": "09/12/2019 09:54:19", "content": "Our models have been not published yet. They are still under research. However, I think you can use this model: \nhttps://github.com/masyagin1998/robin
\n\nOur performance is not much different compared to this. I tried this model and the binarized outputs are quite similar. It should not affect the results.
", "votes": null, "replies": [] } ] }, { "id": 625039, "author_name": "jmartindelasierra", "author_url": "", "post_date": "09/12/2019 16:43:21", "content": "Hi @backaggle, your binarization looks great! Could you give us a clue about your approach (algorithm, strategy, ...)?
\n\nThank you.\nJesús
", "votes": null, "replies": [ { "id": 625363, "author_name": "backaggle", "author_url": "", "post_date": "09/13/2019 01:48:00", "content": "Binarization can be treated as a segmentation problem where you try to classify whether a pixel belongs to the strokes or the background. We use typical segmentation models like: Unet, Deeplab, etc.
One of the problems in this field is the lack of data. The published datasets are totally a few hundred. in both hand-writing and printing type. It is easy to make a ground truth, but it is hard to make a training data because you should make the background looks real. To deal with this problem, we use Cycle-GAN to transfer the background style to the different ground truths. We were successful in this approach.
Interesting topic. Are you going to release anytime soon your algos?
", "votes": null, "replies": [ { "id": 626216, "author_name": "backaggle", "author_url": "", "post_date": "09/14/2019 03:08:56", "content": "Thank for your interest. \nAs soon as we have the result of DIBCO competition this year, we will release our paper.
", "votes": null, "replies": [ { "id": 2466689, "author_name": "amitkchaudhary", "author_url": "", "post_date": "10/04/2023 03:42:02", "content": "Dear Sir, I need complete code of entitled \"Three-stage binarization of color document images based on discrete wavelet transform and generative adversarial networks.\" to complete my master project.
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "624670": "Hi all, \nI wish I have time to take part in this competition because I want to check if our research topic can help in this kind of task. Regarding to document processing, especially hand-written document, we often do `binarization` first. Our group is working on this topic. For better understanding document binarization, please see some samples bellow: \n**Raw image** \n\n\n\n**Binarized image** \n\n\n\n**Raw image** \ntest_0ad567f7.jpg\n\n\n\n\n**Binarized image** \ntest_0ad567f7.jpg \n\n\nThose images are not that bad right? \n\nThey are processed by our pretrained models which using datasets: DIBCO, H-DIBCO, etc. \nEvent those datasets above are different from Kuzushiji dataset, however, I am surprised that the model can perform well. \n\nI upload `binarized dataset` in here: \nhttps://www.kaggle.com/backaggle/kuzushijibin \n\nPlease let me know if the binarized dataset can help you. \nHappy Kaggling", "624703": "Hi, thanks for bringing it up, I’m also curious if binarization will help here. Are binarization models which you used public? As I understand, we won’t be able to the data you posted for submissions if models aren’t public under some appropriate license.", "624708": "Our models have been not published yet. They are still under research. However, I think you can use this model: \nhttps://github.com/masyagin1998/robin \n\nOur performance is not much different compared to this. I tried this model and the binarized outputs are quite similar. It should not affect the results.", "625039": "Hi @backaggle, your binarization looks great! Could you give us a clue about your approach (algorithm, strategy, ...)?\n\nThank you.\nJesús", "625363": "1. Binarization can be treated as a segmentation problem where you try to classify whether a pixel belongs to the strokes or the background. We use typical segmentation models like: Unet, Deeplab, etc. \n\n2. One of the problems in this field is the lack of data. The published datasets are totally a few hundred. in both hand-writing and printing type. It is easy to make a `ground truth`, but it is hard to make a `training data` because you should make the `background` looks real. To deal with this problem, we use Cycle-GAN to `transfer` the background style to the different ground truths. We were successful in this approach.", "626080": "Interesting topic. Are you going to release anytime soon your algos?", "626216": "Thank for your interest. \nAs soon as we have the result of DIBCO competition this year, we will release our paper.", "2466689": "Dear Sir, I need complete code of entitled \"Three-stage binarization of color document images based on discrete wavelet transform and generative adversarial networks.\" to complete my master project." }, "source": "meta" }