kynnemall / dogbreedapp Goto Github PK

A simple web app for identifying dog breeds in colour images. Convolutional neural networks (CNNs) are capable of classifying images using learned feature representations for each class to be identified. State-of-the-art models were trained to determine if images contained a human or dog (MobileNetV3-Small) and then to predict which dog breed the image most resembles. The dog breeds were learned by training on a dataset containing 133 different breeds.

• Clone this repo
• Change directory to the app directory
• Install tensorflow and streamlit using pip
• Enter streamlit run main.py in a terminal. A browser tab containing the web app should open
• Upload a sample image and have fun!

• streamlit was used to build the web app
• tensorflow (version 2.5.0) was used to define model architectures and train models
• scikit-image and scikit-learn were used to implement a local interpretable model-agnostic explanation (LIME) algorithm to better understand how the model makes predictions. This algorithm was implemented according to a tutorial provided by Cristian Arteaga (arteagac), which can be found here

• The dog breed dataset containing over 8000 images of 133 different breeds here
• The human faces dataset (over 13000 images) can be found here

• Transfer learning was used to train models to predict whether an image contains a human or a dog and what dog breed is in the image
• An Xception model was trained to predict dog breeds. It scored 99% and 85% accuracy on the training and validation data, respectively
• MobileNetV3 was trained on equal numbers of dog and human images and achieved 100% accuracy on unseen test images
• MobileNetV3 is currently being trained to distinguish the dog breeds so as to reduce latency compared to the Xception model