CardDetection

CardDetection is a machine learning project designed to detect and identify cards from images. This project utilizes convolutional neural networks (CNNs) to accurately segment cards in the images.
Study can be found here.Link

The dataset used is midv-500

Requirements

• Python 3.9.x
• streamlit==1.36.0
• Cuda 11.6
• torch==2.3.1
• torchvision==0.18.1
• segmentation-models-pytorch==0.3.3
• scikit-learn==1.5.0
• matplotlib==3.9.0
• tqdm==4.66.4

Please install the following dependencies:

pip3 install -r requirements.txt

Preapre the data using

data_preps.ipynb

This will divide the folder into images and masks and create data split by train-->70%, val--> 15%, test-->15%.

How to Train the Model

To train the model, check the config files. These files contain a series of parameters to configure the training process. An example configuration file is config_files/e_unet_res18.json

Here is a sample configuration:

{
  "paths": {
    "IMAGE_PATH_train": "./train/images/",
    "MASK_PATH_train": "./train/masks/",
    "IMAGE_PATH_val": "./val/images/",
    "MASK_PATH_val": "./val/masks/"
  },
  "transforms": {
    "mean": [0.485, 0.456, 0.406],
    "std": [0.229, 0.224, 0.225],
    "train": {
      "resize": [256, 256],
      "random_horizontal_flip": true,
      "random_vertical_flip": true
    },
    "val": {
      "resize": [256, 256],
      "random_horizontal_flip": false
    }
  },
  "training": {
    "batch_size": 32,
    "max_lr": 0.001,
    "epochs": 15,
    "weight_decay": 0.0001,
    "l2_reg_bool": false,
    "lambda_l2": 0.001
  },
  "model": {
    "model_name": "unet",
    "encoder_name": "resnet18", 
    "classes": 2,
    "activation": "softmax",
    "encoder_depth": 5,
    "decoder_channels": [256, 128, 64, 32, 16],
    "resume_path": ""
  }
}

To train the model, run the following command:

python3 main.py config_files/e_FPN_mobilenet_v2.json

For Inference or Web-Based Tool

Note: The inference is run using streamlit. Prepare the data in mask and images folders and replace the paths in the script accordingly.

To run the inference, use the command:

streamlit run inference.py