Aric Prieve
This is my final project for CSE 490g Autumn 2022 at the University of Washington. The code for this project is here.
My goal for this project was to train a classifier to predict which national park an image was taken from. As data is at the core for successful deep learning I wanted to experiment both with training a convolutional network on new data and collecting that data to see the effects of a new type of dataset. I wanted to learn how difficult gather and assembling a dataset for computer vision was and if this type of data collection could yield something interesting/meaingful. In this project I did just that on the national parks in the US. I learned a lot about data collection and thought I had some decent results in the end.
I will create a datset of images of each of the National Parks in the US with about 350 samples each. These images will be scraped from the internet and will need to be standarized. This standardization includes normalization and sizing (256x256 images were used). I then will train a convolutional neural network of a few different types to compare performance both of the models on the dataset and whether this dataset needs improvement.
The first portion of this project was gathering the data. The code for this step in get_images.ipynb. I first downloaded a csv of the national parks to serve as the labels for the dataset. Next was gathering the images. I chose to download 500 images of each national park using web scraping on google images with the search query "[national park] National Park Landscape". Initally I found this difficult as much of the APIs in python for this type of task were depracted. I then found a perfectly functioning API for this task, but would use Bing as the search engine. I chose to use this which worked well until the realization that Bing is not a good search engine. Many of the downloaded pictures would not actually refer to search query and would show Yosemite (probably a popular national park) as the image. I then reran this task using simple_image_download which was less reliable in terms of downloads (would download a number of images from 320 - 400), but resulted in much higher quality images that actually matched the query.
The downloaded images still needed to be in a standard size. Most images were much higher resoltion than needed (or desired) and not a consistent aspect ratio so I downsized them, also in python, to 256x256 images. Now with around 350 images for each national park I was ready to use them for training a neural network with an 80 / 20 train / test split.
I then created another python notebook similar to homework 1 from this quarter. I created a custom torch dataset where I normalized the images, batched, and labeled the images similar to homework 2. I started by defining my model similar to my result from homework 1 which was trained on TinyImageNet but adjusted and tinkered with different model architecture and hyperparameters. Finally I plotted my results across those changes and showed the weaknesses of the model (or the dataset) with a confusion matrix.
I chose to analyze my results for the model similarly to the way we have done in other homework assignments by plotting the training/test and loss/accuracy curves. I also wanted to show a confusion matrix in hopes to see which national parks might be confused with one another.
My model did surprsingly well. My final accuracy was 99% with the model I chose to use (all details on model specs in the code). Below are the standard performance plots. These results lead me to believe that this problem for the neural network architecture was unexpectable easy. I think this is the case becuase many of the images were quite repetetive, as you would get from a google search. So Although the images might not have the exact same for in the test and train, many of the images are extremely similar as for most national parks there are the few shots that appear often. If I were to continue this I would want to expand the size of the dataset and widen the assortment of images. Like I said, much of the images are repetetive and of the same scene, so I would want to find more nuanced pictures in the hope most pictures randomly taken in the park could be classified.
Link here
The main resources I used were simple_image_download the API and referencing/using the heavily the code from the python notebook assignments this quarter.