[Feat] Addtional information about nanoparticles and pores

[yoonjongyeon]

What

Enhance U-Net's performance by adding information about nanoparticles and pores. The information we are talking about here is the number of nanoparticles and pores inside the image and the number of pixels each of them contains.

Why

Using this additional information might enhance the performance (e.g. dice score) of the U-Net.

TODO

• [x] Use BFS algorithm to get the information we want to get and save them as pickle.
• [x] Update .csv files and dataloader.

[yoonjongyeon]

https://github.com/yehyunsuh/TEM-Image-Segmentation/blob/da7b2cc494020140951801ba54948149c78d625e/bfs.py#L4-L35

BFS code for getting the number of nanoparticles and pores inside the image and the number of pixels each of them contains.

[Input]

• visited: empty image (must be the same dimension and size as the original image)
• n: row of the original image
• m: column of the original image
• grid: original image (the image we want to get information from)

[Output]

• visited: the result of the BFS algorithms (0 for background, 255 for nanoparticle or pores)
• count: a list that contains the number of pixels in each nanoparticles or pores

[yoonjongyeon]

https://github.com/yehyunsuh/TEM-Image-Segmentation/blob/da7b2cc494020140951801ba54948149c78d625e/bfs.py#L38-L65

Code that makes the list's size the same to put them in dataloader or to compare them with others.

[Pore]

• Made the list size to the maximum number of pores in a single image. (zero-padded when smaller and cut them when bigger)

[Nanoparticle]

• Made the list size to the maximum number of nanoparticle (2) in a single image. (zero-padded when smaller and cut them when bigger)

[yoonjongyeon]

https://github.com/yehyunsuh/TEM-Image-Segmentation/blob/087c8222a5b904071f8e0e452955a476039004ec/dataset.py#L58-L63

Updated dataset to return additional information.

• num_unit: number of nanoparticles or pores in an image (int)
• num_pixel: number of pixels that each nanoparticles or pores contains (list)

[yoonjongyeon]

https://github.com/yehyunsuh/TEM-Image-Segmentation/blob/f4d4e0551bba137bd3b3c9c04a1db0039398e5f9/train.py#L44-L77

Added RMSE between the difference of unit numbers and pixels in label and prediction. This value is added to loss when using distribution loss.