legitimization-experiments

Neural network training graphs and inference results to go along with my thesis "Legitimizing monitored subjects using image recognition techniques" The topic of the thesis concerns training and/or utilizing a set of prediction models including:

• an image classifier for detecting the LED of breathalyzer devices
• a custom object detector for detecting breathalyzers in a picture
• face detection and recognition tools

What's to see here?

A bunch of Jupyter notebooks and tensorboard graphs:

• color-inception-v3/scripts.ipynb - Notebook and tensorboard logs for re-training InceptionV3 architecture in Keras.
• sagemaker-detections/detector-train.pynb - Notebook and script for re-training and tuning the keras-retinanet detector in Keras.
• multi_check_test.ipynb - Inference on a multi criteria check: detect faces, device, recognize subjects and verify device is used by the correct person.

Launch it in jupyter

build.sh, run.sh - execute to build & run Jupyter docker image with dependencies.

Previews

Training color classifier (from TensorBoard):

Inference on multi criteria check (from Jupyter):

Adadelta, RMSProp optimizer graphs in hyperparameter tuning (from SageMaker):

Thesis: Legitimizing monitored subjects using image recognition techniques

I will add the link to my thesis as soon as it is publicly available on the university's server.