This repository contains the projects for the Artificial Neural Networks and Deep Learning Course. For all these projects we used Tensorflow 2.1.0 (all requirements for each competition are listed in the corresponding requirements.txt file).

Authors: Francesco Alongi and Gerlando Re

The first project was an Object Recognition challenge in which we had to build a CNN classifier. The dataset was composed of images representing 20 different types of objects (airplanes, bulbs...) also with an unbalance in the number of samples.

The second project had the purpose of getting acquainted with the Image Segmentation problem, which consists in selecting the pixel space within the image in which a specific object "lies". For this network we did some literature research in order to study effective models to solve the problems (U-net etc.). For this challenge we used a dataset containing satellite images of buildings and road, where the recognition targets are the building shapes.

The third and last challenge was a Visual Question Answering challenge, with the purpose of building a model that could intertwine both a CNN and an RNN. The challenge consisted of answering to some questions about a particular image, selecting the answer among 13 possible predefined answers. For this challenge a simpler version of the Stanford CLEVR dataset was used.
Also for this challenge we had to explore the literature in order to find a model that could adapt to our problem and could give us a good performance. Also the special side of the challenge was to mix a network for the feature extraction of the image with a recurrent neural network for text prediction.