Thesis-NoiseDetection-CNN

Noise detection with CNN

Requirements

git clone --recursive https://github.com/prise-3d/Thesis-NoiseDetection-CNN.git

pip install -r requirements.txt

Project structure

Code architecture description

• modules/*: contains all modules usefull for the whole project (such as configuration variables)
• analysis/*: contains all jupyter notebook used for analysis during thesis
• generate/*: contains python scripts for generate data from scenes (described later)
• prediction/*: all python scripts for predict new threshold from computed models
• simulation/*: contains all bash scripts used for run simulation from models
• display/*: contains all python scripts used for display Scene information (such as Singular values...)
• run/*: bash scripts to run few step at once :
  • generate custom dataset
  • train model
  • keep model performance
  • run simulation (if necessary)
• others/*: folders which contains others scripts such as script for getting performance of model on specific scene and write it into Mardown file.
• custom_config.py: override the main configuration project of modules/config/global_config.py
• train_model.py: script which is used to run specific model available.
• prediction_model.py: script which is used to run specific model with data in order to predict.

Generated data directories:

• data/*: folder which will contain all generated .train & .test files in order to train model.
• saved_models/*: all scikit learn or keras models saved.
• models_info/*: all markdown files generated to get quick information about model performance and prediction obtained after running run/runAll_*.sh script.
• results/: This folder contains model_comparisons.csv file used for store models performance.

How to use

Generate reconstructed data from specific method of reconstruction (run only once time or clean data folder before):

python generate/generate_reconstructed_data.py -h

Generate custom dataset from one reconstructed method or multiples (implemented later)

python generate/generate_dataset.py -h

Reconstruction parameter (--params)

List of expected parameter by reconstruction method:

• svd_reconstruction: Singular Values Decomposition
  • Param definition: interval data used for reconstruction (begin, end)
  • Example: "100, 200"
• ipca_reconstruction: Iterative Principal Component Analysis
  • Param definition: number of components used for compression and batch size
  • Example: "30, 35"
• fast_ica_reconstruction: Fast Iterative Component Analysis
  • Param definition: number of components used for compression
  • Example: "50"
• diff_filter: Bilateral diff filter
  • Param definition: window size expected
  • Example: "5, 5"
• sobel_based_filter Sobel based filter
  • Param definition: K window size and pixel limite to remove
  • Example: "3, 30"
• static Use static file to manage (such as z-buffer, normals card...)
  • Param definition: Name of image of scene need to be in {sceneName}/static/xxxx.png
  • Example: "img.png"

Example:

python generate/generate_dataset_sequence_file.py --output data/output_data_filename --folder <generated_data_folder> --features "svd_reconstruction, ipca_reconstruction, fast_ica_reconstruction" --params "100, 200 :: 50, 10 :: 50" --sequence 5 --size "100, 100" --selected_zones <zones_files.csv>

Then, train model using your custom dataset:

python train_lstm_model.py --train data/custom_dataset.train --test data/custom_dataset.test --chanels "1,3,3" --epochs 30 --batch_size 64 --seq_norm 1 --output output_model_name

Predict image using model

Now we have a model trained, we can use it with an image as input:

python prediction/predict_noisy_image.py --image path/to/image.png --model saved_models/xxxxxx.json --features 'svd_reconstruction' --params '100, 200'

• features: feature choices need to be one of the listed above.

The model will return only 0 or 1:

• 1 means noisy image is detected.
• 0 means image seem to be not noisy.

Simulate model on scene

All scripts named prediction/predict_seuil_expe*.py are used to simulate model prediction during rendering process.

Once you have simulation done. Checkout your threshold_map/%MODEL_NAME%/simulation_curves_zones_*/ folder and use it with help of display_simulation_curves.py script.

License

MIT