The group identified gaps within the "Foodie" market to create a business solution using machine learning. In this case, we sought to provide online Foodies access to recipes quickly from the simple use of images that can be screenshot from social media applications.
Implemented computer vision techniques using TensorFlow-Keras package in Python. Including but not limited to: cleaning the images, creating data loaders for the rgb image arrays, implementing data augmentation techniques, and, finally, modeling using Convolutional Neural Networks.
Tuned the following CNNs: ResNet50, ResNet101, DenseNet201, MobileNetV3, InceptionV3, ConvNeXtBase, & EfficientNetB0 to the Food-101 dataset.
The team then built a FastAPI application using the preferred model, MobileNetV3, due to its overall performance (77.8% accuracy) in addition to its less intensive computational costs as the smallest CNN of the group. The computation costs were vital in the decision making as the team wanted users to have a fast interface to make predictions with.
Committed high level data cleaning on review data to prep for modeling that included: removal of stop words, tokenization, removal of punctuation, and removal of uppercase letters.
Scraped the review data from Amazon HTML pages using Selenium in python.
Constructed sentiment model with 83.2% accuracy on test set (Support Vector Classifier).
FastAPI model deployment with HTML/CSS endpoint construction and formatting.
