Abstract

Object detection or localization gradually progresses from coarse to fine digital image inference. It
provides the classes of the image objects and the location of the image objects that have been categorized.
The location is provided in the shape of bounding boxes or centroids. Semantic segmentation provides fine
inference by indicating labels for every pixel in the input image. Each pixel is labelled according to the
object class within which it is surrounded. Moreover, instance segmentation gives different sets of pixels of
interest for separate instances of objects. Thus, instance segmentation is defined as solving the problem of
object detection and semantic segmentation simultaneously.
This thesis is introducing a new self-supervised framework for instance segmentation and background
removal. We make use of a state-of-art self-supervised method called bootstrap your own latent for our
pretraining section and then by transfer learning, we transfer the representation learnt in this pretraining
to downstream task which is instance segmentation using Mask R-CNN. Two other existing state-of-art
methods of instance segmentation, namely, instance segmentation scheme using Mask R-CNN along with
random initial weights and that with ImageNet initial weights, respectively are implemented and compared
to our proposed framework. Comparing our method with those known methods in instance segmentation and
background removal, we find out that self-supervised learning outperforms both methods despite using no
labelled data in pretraining. Our experimental results indicate that the proposed framework outperforms the
instance segmentation and background removal using ImageNet initial weights and random initial weights
by 0.866% and 14.06% ,respectively, in average precision (AP). Moreover, the proposed self-supervised
framework has been designed to minimize the computations and the need for annotated datasets. This
design demonstrates that the proposed system can perform high-quality processing at a meagre computation cost for many applications. Hence, instance segmentation using self-supervised techniques is a practical
approach to lowering the barrier of computation resource requirements and labelled datasets to make them
more implementable and applicable to the general public.