Hello guys, I want to perform segmentation on my images so that I will train them using an Unet model ….so I want to create a function to mask theme images…i have already converted my images to png …so I want the function when I call the function should mask the image and visualize the original and the masked image…example of usage of the function.
\nmask = create_mask(train_df,sample_img)
This is example of the function I used to mask the images but didn't work pretty well.
\n`def create_mask(df: pd.DataFrame, image_name: str='img.png', shape: tuple=(350, 525)):
masks = np.zeros((shape[0], shape[1], 1), dtype=np.float32)\n\ndf = df[df[\"img_path\"] == image_name]\n\nfor idx, im_name in enumerate(df[\"img_path\"].values):\n\n label = df.loc[df[\"img_path\"] == im_name, \"cancer\"].values[0]\n\n mask = cv2.imread(f\"/kaggle/working/output/train/{label}/{im_name}\")\n\n if mask is None:\n\n continue\n\n masks[:, :, 0] = mask[:, :, 0] / 255\n\nreturn masks\n `\nthe masks output was dark…I am trying to use the techniques which was used here Image Segmentation
", "messages": [ { "id": 2142751, "postDate": "2023-02-13T18:50:52.887Z", "content": "Hey Denis, the code isn't showing up properly, but I don't see how that could work. I saw @paulbacher processing notebook. This notebook is excellent btw, and thank you Paul for posting it. it had a masking technique. He used it to help with cropping but I think it works really well.
\nhttps://www.kaggle.com/code/paulbacher/custom-preprocessor-rsna-breast-cancer
\nEssentially this would be the code where img is your numpy of image shape (h,w)
\nthreshold = 0
\nbin_img = (img > threshold).astype(np.uint8)
\ncontours, _ = cv2.findContours(bin_img, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
\ncontour = max(contours, key=cv2.contourArea)
\nmask = np.zeros(img.shape, np.uint8)
\ncv2.drawContours(mask, [contour], -1, 255, cv2.FILLED)
Some key piece might be missing here. 🤔 Have you predicted segmentation masks with some method (e.g., CNN)? And if so, are those masks located in \"/kaggle/working/output/train/{label}/{im_name}\"? The RSNA competition data does not include pixel-wise segmentation masks of the lesions, but you indeed could estimate those from the data with suitable technology.
Hello guys, I want to perform segmentation on my images so that I will train them using an Unet model ….so I want to create a function to mask theme images…i have already converted my images to png …so I want the function when I call the function should mask the image and visualize the original and the masked image…example of usage of the function.
\nmask = create_mask(train_df,sample_img)
This is example of the function I used to mask the images but didn't work pretty well.
\n`def create_mask(df: pd.DataFrame, image_name: str='img.png', shape: tuple=(350, 525)):
masks = np.zeros((shape[0], shape[1], 1), dtype=np.float32)\n\ndf = df[df[\"img_path\"] == image_name]\n\nfor idx, im_name in enumerate(df[\"img_path\"].values):\n\n label = df.loc[df[\"img_path\"] == im_name, \"cancer\"].values[0]\n\n mask = cv2.imread(f\"/kaggle/working/output/train/{label}/{im_name}\")\n\n if mask is None:\n\n continue\n\n masks[:, :, 0] = mask[:, :, 0] / 255\n\nreturn masks\n `\nthe masks output was dark…I am trying to use the techniques which was used here Image Segmentation
", "rawMarkdown": "Hello guys, I want to perform segmentation on my images so that I will train them using an Unet model ....so I want to create a function to mask theme images...i have already converted my images to png ...so I want the function when I call the function should mask the image and visualize the original and the masked image...example of usage of the function.\n`mask = create_mask(train_df,sample_img)`\n\nThis is example of the function I used to mask the images but didn't work pretty well.\n`def create_mask(df: pd.DataFrame, image_name: str='img.png', shape: tuple=(350, 525)):\n\n masks = np.zeros((shape[0], shape[1], 1), dtype=np.float32)\n\n df = df[df[\"img_path\"] == image_name]\n\n for idx, im_name in enumerate(df[\"img_path\"].values):\n\n label = df.loc[df[\"img_path\"] == im_name, \"cancer\"].values[0]\n\n mask = cv2.imread(f\"/kaggle/working/output/train/{label}/{im_name}\")\n\n if mask is None:\n\n continue\n\n masks[:, :, 0] = mask[:, :, 0] / 255\n\n return masks\n `\n\nthe masks output was dark...I am trying to use the techniques which was used here [Image Segmentation](https://github.com/sneddy/pneumothorax-segmentation/blob/master/unet_pipeline/utils/prepare_png.py)" } ], "comments": [ { "id": 2142751, "author_name": "Michael Bolton", "author_url": "", "post_date": "2023-02-13T18:50:52.887000", "content": "Hey Denis, the code isn't showing up properly, but I don't see how that could work. I saw @paulbacher processing notebook. This notebook is excellent btw, and thank you Paul for posting it. it had a masking technique. He used it to help with cropping but I think it works really well.
\nhttps://www.kaggle.com/code/paulbacher/custom-preprocessor-rsna-breast-cancer
\nEssentially this would be the code where img is your numpy of image shape (h,w)
\nthreshold = 0
\nbin_img = (img > threshold).astype(np.uint8)
\ncontours, _ = cv2.findContours(bin_img, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
\ncontour = max(contours, key=cv2.contourArea)
\nmask = np.zeros(img.shape, np.uint8)
\ncv2.drawContours(mask, [contour], -1, 255, cv2.FILLED)
Some key piece might be missing here. 🤔 Have you predicted segmentation masks with some method (e.g., CNN)? And if so, are those masks located in \"/kaggle/working/output/train/{label}/{im_name}\"? The RSNA competition data does not include pixel-wise segmentation masks of the lesions, but you indeed could estimate those from the data with suitable technology.