Image Classification

Inspiration

We were immediately drawn to this track for the place it holds in AI's future, and also to increase our knowledge on the subject. Our task was not to create new, but rather compare and evaluate existing image classification models. We were inspired by those that have been known to do it well, and that is open-source (). Based on the existing source code, we were able to thoroughly understand how each deep learning and precisely image classification achieved, and how altering just the last few layers could primarily ease the training process. The GDIT workshop was also instrumental in clarifying the problem and showing us the extent/target of Image Classification, and why it is increasingly important to optimize this for the federal sector.

What it does

The project compares VGG16 and VGG19 pre-trained models in image classification. It divided into two main parts: Predict image classification based on the pre-trained model: We used VGG16 and VGG19, and based on the testing, VGG16 is a better model.

Edit the pre-trained model to predict images based on the nine classifications in the images folder: Also, the VGG16 performs better in the edited version.

How we built it

• We used the Python and Keras library. For more details, please check the GitHub link below.

Challenges we ran into

• The way of using pre-trained models was not clear initially, and we had to take some time to understand what was expected
• At the beginning, we went the wrong way and mainly focused on how to use the sample images to train our own models. It indeed took us a long time to understand others' programs and how to make them work for us appropriately. The codes we studied were basically about image binary classification, which means it can only deal with the question like classifying cat and dog. But as time went on and we had a better understanding about it, we could implement multiple images in our program even though the accuracy was not that good. Although it was not what we were supposed to do, we still learned a lot from the experience.

Accomplishments that we're proud of

• We are proud that we finished earlier than expected, and explored in a different direction than the track intended; we did all we could given the size of the data set.

What we learned

• We understood the fundamentals of machine learning and deep learning, became familiar with Keras and re- and pre-training, and explored the time and accuracy costs of having more layers. And though we didn't use it, we learned the basics and importance of TensorFlow, as well as other important facets of NN.

What's next for Image Classification

• For the second part of the project, we need to train the model with a bigger dataset, and fine-tune the number of layers to yield maximum accuracy. We also need to evaluate why certain images were predicted so incorrectly (women were said to be hairspray), and why VGG16 seemed to perform better than VGG19.

Built With

• jupyter-notebook
• keras
• python