Contains the training, saving, and evaluation of a deep learning model trained to predict the presence of internal water pipe defects from ultrasound data, made using TensorFlow and Keras libraries in Python. The model's predictions is outputted in float32 format ranging as a number from 0 to 1; 0 being no defect is present and 1 being there is a defect present in the image inputted. The closer the prediction is to 1, the higher the model's confidence that the defect is present.
Model specifications are that it used ResNet50 as a base model with additional layers added to the top and bottom of the neural network, SGD was used as the optimizer as it is non-adapative (did not cause overfitting like the Adam optimizer), and that the .hdf5 model accepted 256x128 png images while the .pb model accepted one encoded string in .b64 format that once decoded could be any size of data.
Also contains Python scripts load in model checkpoint files (.hdf5) to receive their predictions from an inputted dataset then plots the improvement from an old model versus a new model, a text file containing Docker commands to host the TensorFlow serving of a model onto Docker, Python scripts that alter existing 256x128 or 500x312 images by changing their contrast or vertically/horizontally shifting, and a MATLAB function that interacts with a GUI to create new nominal images based off existing project's data.