Implementation of Deep Learning Techniques for Image Segmentation and Object Recognition

Abstract

Object Recognition and Image segmentation plays an important role in a pre-processing phase of an image. Only objective of segmenting an image is to partition the input file into region of interest resulting in a more refined analysis of one or multiple regions. Every object in the supplied image is pinpointed and located using object recognition. Currently Faster R-CNN (Regions with Convolutional Neural Network) is used for image acquisition and localization. Image classification merely categorises items from the input, whereas semantic segmentation recognises the same and classifies it as a single instance from the same class. However, we cannot Mget the count of similar objects present in that particular instance. Whereas, instance segmentation is more precise and draws in when there are multiple objects to deal with. The main distinction is that the identified item is masked, and the pixels linked with it are all assigned the same colour. Because many objects of the same class are recognised as distinct entities, each entity is masked with a separate colour. This paper deals with instance image segmentation. In this paper, with the help of Mask R-CNN we predict both bounding boxes and pixel wise segmentation masks for every object. So that in the provided input image we can segment the foreground and the background at a much granular level. We are also addressing the task of object recognition using deep learing techniques and python. Our proposed system results in better segmented images, because Mask R-CNN masks every object and produces more granular result.

Objects belonging to the same class are masked with different colors which makes it easier to distinguish similar objects.

Keywords: R-CNN, Object Detection, Instance Segmentation, Region of Interest, Convolutional Neural Network