This GitHub repository contains the code we developed for our final project at Centrale Marseille.
Author: Cléa Han, Yanis Labeyrie, Adrien Zabban
This project focuses on the analysis of blood traces within the scope of the work of forensic expert Philippe Esperança. The objective is to develop an artificial intelligence to assist and streamline the analysis of crime scenes involving blood.
Philippe Esperança's work on blood trace analysis has resulted in the classification of these traces into 18 distinct classes, listed in the table below. Our goal is to create a deep learning model capable of predicting the class for an image of a blood trace.
| Classes | Classes |
|---|---|
| Passive Traces Models | Droplet Models |
| Contact Transfer Model | Sliding Transfer Model |
| Contact Alteration Model | Sliding Alteration Model |
| Accumulation Model | Flow Model |
| Volume Fall Model | Propelled Blood Model |
| Ejection Model | Impacted Volume Model |
| Impregnation Model | Interruption Zone Model |
| Impact Model | Impact Focal Point Model |
| Gravitational Trace Model | Expired Blood Model |
To run the code you need python (We use python 3.10.11) and packages that is indicate in requirements.txt. You can run the following code to install all packages in the correct versions:
pip install -r requirements.txtTo perform an inference, execute the code run_infer.py with the following arguments:
| Command | Description | Default Value |
|---|---|---|
| -d | Path to the folder containing the images to predict | |
| -m | Path to the model to use | logs/retrain_resnet_allw_img256_2 |
| -o | Path where the results will be saved (creating this folder if necessary) | Subfolder "inference_results" in the data directory |
| -s | Option to generate saliency map (true or false) | true |
You can use python run_infer.py -h for this documentation. Example of code execution:
python run_infer.py -d data/images_to_infer -m logs/retrain_resnet_allw_img256_2 -o data/output -s falseYou can also perform an inference with streamlit by using run_app.bat if you use Windows, or run this commende line:
streamlit run streamlit/streamlit_app.pyIn the file config.yaml, you can specify all details about your model such as data augmentation, model architecture, hyperparameters, batch size, learning rate, etc. Once your training is completed, a copy of this configuration will be saved in the logs directory along with the model weights, its performance metrics, and learning curves.
For the code to function properly, the data for training, validation, or testing must be organized into subfolders named respectively train_256, val_256, and test_256, where 256 represents the image size (change the names of the subfolders accordingly). These subfolders should be grouped within a single directory, which you will specify in the file config.yaml under data.path.
For example:
data: # data parameters
path: data/data_labo # path to the data
real_data_path: data/data_real # path to the real data
image_size: 256 # size of the imagesIn this example, the subfolders train_256, val_256, and test_256 containing laboratory images and real data respectively are within the directory data/data_labo and data/data_real.
To perform training, validation, testing, or even launch random search, you should use the file main.py. You can run python main.py -h for more information.
Here are examples of launching training with the default configuration from config:
python main.py --mode trainAnd an example of testing the resnet_allw_img256_2 model without measuring the saliency map metrics:
python main.py --mode test --path logs/resnet_allw_img256_2 --run_saliency_metrics falseTo conduct a random search or a grid search to find the best hyperparameters, you need to create a file named search.yaml in the config folder with the parameters you want to test. For example, you can test finding the best learning rates and the alpha parameter for the adversarial model. Copy the following example into search.yaml:
learning:
learning_rate: [0.01, 0.005, 0.001, 0.005, 0.0001]
adv:
learning_rate_adversary: [0.01, 0.005, 0.001, 0.005, 0.0001]
alpha: [0.001, 0.1, 0.5, 1, 2, 5, 10, 100]Then, to launch 20 training runs:
python main.py --mode random_search --num_run 20If you want to test all combinations, use grid search instead of random search. This will create a directory in logs containing all your experiments. You will also have a summary table of the performance for each experiment in a CSV file.
Our strategy is described in the report (which is in French) that we invite you to read to better understand what we have done. Unfortunately, we had to remove the images from the report because they are confidential. You can use the table below, which provides the correspondence between the names given in this report and the names given in this repository.
| Models name in the report | Model name in this repositoy |
|---|---|
| LP ResNet | resnet_img256_0 |
| FT LP ResNet | retrain_resnet_img256_0 |
| AWL ResNet | resnet_allw_img256_2 |
| FT AWL ResNet | retrain_resnet_allw_img256_2 |
| Adversarial | adv_img256_1 |
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent