A convolutional neural network that counts the number of fingers held up in front of a webcam.
This network was created for the course Begleitseminar: Product Identification from Images and Video.
data_collection.py: The python script to create a dataset of finger images.data_loader.py: Loads the data into memory and creates batches for training.keras: Folder where the neural network architectures are.
data_collection.py will capture any number of images, label them, and create a file structure to store them. You can take any number of images at any time, and change which label you want to put on the images, on the fly. Pressing q will quit if necessary. Press h for help. After creating the images, the script will move each image into its respective folder (named after the images' label).
The file structure used is five folders, each named after their respective label, placed in the root directory. This structure was chosen because it allows easy access to check the data quickly during acquisition (i.e., no need to go click more than necessary).
data_loader.py will step through the folder structure and load each image into memory. It will then split it into train (60%), validation (20%), and test (20%) datasets. It also contains a method to create a generator object to step through training/validation/test batches.
finger_resnet.py is a ResNet implementation intended to be trained on the dataset. Currently, it runs into dimensionality issues where the channels explode into the 1000s.