Image segmentation and pre-processing for electronic waste identification - Using OpenCV to compare different techniques for object extraction and rotation

Abstract

This is a master thesis which compares several methods for foreground segmentation and object rotation. By object rotation, it is meant in this thesis that for a given object, having images with different angles as input should ideally always output images with the object in the same angle. Finally, it is tested how this combined can make the task of object recognition easier by running the algorithms for the two tasks successively as an image pre-processing stage. It is shown that making pixel-wise background segmentation by comparing the input image to an averaged background image works well for the segmentation task, and that encapsulating the object in a minimum bounding rectangle and rotating it with the angle of the bounding box can work well for the rotation task. It is also shown that using these algorithms combined as a pre-processing stage to an object classifier may be a way of making the classification easier.