/Muscle-enhancing-via-Autoencoder

This is just a funny project that we want to see AutoEncoder (AE) can actually work to enhance the features we want. We will start to improve...

Primary LanguageJupyter Notebook

Muscle-enhancing-based_on-Autoencoder

1.Discription:

  1. This is just a funny project that we want to see AutoEncoder (AE) can actually work on the some features. We will start to improve...
  2. Origial image and results for iterating 1~5-times.

image

2.Future work:

  1. Add more data to train
  2. Higher resolution
  3. Imporve preprocess (Inverse_Muscle_filter.ipynb): denoise process/enhance not incude face
  4. postprocess: brightness fixed/denoise process.

3.Usage:

  1. Downloads pre-trained model and put in the folder.
  2. Open "Pred.ipynb" .
  3. Input the image name you wwant to test.
  4. Run the whole code.

4.Training by yourself

  1. Downloads a lot of muscle image from internet (Since the copy right problem, I cannot share my dataset with you). The images type can be jpg/png/jfif/... . Notice: The more visible the muscle lines are in the images, the better. In our case, we have 204 images now.
  2. Download repository-skin detector-1 and skin detector-2. Then, put their with code.
  3. Create 2 folders: before_filtering/after_filtering. Put the downloaded images in to "after_filtering" folder. Also create 2 empty folders: before_filtering_rm_bg/after_filtering_rm_bg which will load images from Inverse_Muscle_filter.ipynb
  4. Run Inverse_Muscle_filter.ipynb.
  5. Open training.ipynb and run the code with suitable epochs.

5. Update History:

  1. [2021/12/21]
  2. Add 100+ image into dataset and remove gray style image (skin detector will not work).
  3. Modify preprocess to can output double images (weakening/original version with/without background). Original code just can output weakening/original version with background.
  4. Change training/prediction shape from (224,224,3) to (448,448,3).
  5. Training a model with size:(448,448,3) and put the new model in pre-trained model
  6. Training condition as shown as following: (x: epoch num, y: mse error of pixels)

image