Scale-invariant image segmentation using machine learning

Abstract

The increased application of segmentation requires more robust machine learning algorithms that can handle variations of the input. The areas of robotics, self-driving cars, automated drone delivery systems, and Speed Enforcement Cameras (SEC) all rely on accurate predictions from machine learning algorithms. CNNs are the current state of the art in image recognition which has led to their increased application in areas of classification, object detection, and segmentation. However, scale invariance poses a significant issue for CNNs; fixed kernel size hinders the prediction accuracy of the network for objects of varying sizes. This research introduces a training methodology and image pre-processing techniques which makes these networks more robust or invariant to the changes in the size of the object.