This project implements a deep learning pipeline using TensorFlow/Keras to continuously fine-tune and train a model for recognizing clothing brands based on labeled image data. The script loads the best available pre-trained model, modifies its final layers for classification, and applies data augmentation to enhance training. It includes automatic stopping criteria based on accuracy thresholds (85% validation accuracy) and detects training plateaus to prevent unnecessary computations. The training process dynamically adjusts learning rates, class weights, and layer unfreezing for optimal fine-tuning. Users can interact with the script to manually override training conditions or continue improving the model. The trained models, along with metadata (brand mappings, training history), are saved for future use. This project started with the utilizeation of OCR but due to the nature of format changing and difficult to read labels I switch to classifiying images using convolutional neural networks.
The project idea is primarily focused on the idea of thrifting or second had shopping to enable the user to see what kind of discounted deals they are getting on the desired item. Project scope is fairly limited as training models such as this required hundreds of images, and with clothing labels that can be upwards to double digit styles that also require more images. These images are taken by me because sorting through places like Google Images is too unpredicatable and actually more time consuming than physically going and taking shots. It is also limited in the sense that I can only focus on small sections of things such as mens apparel or books genres that I know I would enjoy, simply put because that is what I shop for and collecting data on everything would be too large of a task.
Launching a functional app that allows users to scan clothings labels, books, and anything that I can think would be the end product and this could also allow for further data collection that I otherwise would be unable to accomplish myself.
