Building Recognition of university buildings using a pre-trained ResNet50 archetecture

The purpose of this project is to create a model that can identify the building shown in an image taken by a user on the UNCC campus. The reason behind this is that there's currently no dataset available that has images of all the buildings on the UNCC campus. This project is not only a fun way to learn about CNNs, but it's also useful as the model. If accurate, this project could be used on other campuses or in large cities where multiple buildings need to be recognized. The project is limited to the UNCC campus, but the possibilities for its use in the future are endless. Imagine being able to use it for self-guided tours or even navigation. The best part? We get to create the dataset ourselves, making the project even more meaningful to the UNCC community.

For this project, we used a pretrained ResNet50 architecture as we determined from our research that this model would provide us with the best results for our use case. Despite what was found in our research, we still tested and examined the results of using other model types, such as using a CNN from scratch, or using different pretrained models. It was found that using ResNet50 would be the best for us. The hyperparameters used included a learning rate of 0.0001, 15 epochs, an image size of (224 x 224), a horizontal flip data augmentation technique, and data normalization that was zero-centered. The loss function used was cross-entropy. The Adam optimizer was chosen as it is computationally efficient and better suited for non-linear applications than the alternative, Stochastic Gradient Descent (SGD).

The resulting model provided us with an accuracy of around 85%