Image Angle Correction Using Neural Networks

Photo by Ant Rozetsky on Unsplash

How cool it would be if we could read the angle of images, just using Neural Nets. It can be used to stabilize the videos in real-time or offline. It can also be used to replace physical sensors used for measuring angles.

It sounds fancy, so I decided to pick this project.

Note

This article is written by keeping in mind that you are already familiar with Neural networks. I am using the Keras framework for this project. Let’s get started.

Chapter 1: Data Curation

Photo by Malek Dridi on Unsplash

Data curation was not that hard. I took a dataset containing 40K outdoor images and rotated them at random angles between -30 to 30 degrees. I resized the final images to 128*128 pixels. I kept images resolution small to make the training process faster.

Labels for images are nothing but the angles from -30 to 30 degrees. Following is the link to the dataset.