Aircraft Identification System

• Developed a computer vision system that identifies aircraft models based on a picture, written in Python with PyTorch. • The models were trained and evaluated with Google Colab GPUs. • The system was deployed with an AWS Sagemaker Endpoint, using AWS API Gateway, AWS Lambda Function and AWS S3. • Used transfer learning techniques. Several pre-trained convolutional network models were evaluated, with the best results being achieved with VGG-19. • Several techniques of data augmentation were used to overcome the low number of training examples per category (only 100 examples of each category in the dataset used). • Reached ~60% Top-1 accuracy and ~86% Top-5 accuracy across 70 categories of previously unseen test set.