An improved semi-supervised learning framework for Image semantic segmentation

Abstract

Traditional supervised learning methods depend heavily on labeled data, which is both costly and time-intensive to acquire. Self-supervised learning approaches present a promising alternative to supervised learning, enabling the utilization of unlabeled data. Thus, this research aims to build an advanced semi-supervised semantic segmentation model that strikes a balance between selfsupervised and fully supervised paradigms for visual perception applications in an autonomous driving environment. In this direction, the thesis is structured into three distinct phases, beginning with self-supervised image classification and progressing toward bi-level image segmentation, ultimately culminating in the development of an advanced semantic segmentation model. Initially, this research employs a simple contrastive learning framework (SimCLR) to classify medical images, specifically focusing on monkeypox diagnosis from skin lesion images, while integrating a federated learning (FL) framework to ensure data privacy. Monkeypox classification is a simple binary classification task and the dataset found for this problem, in this thesis, is very manageable on the computational resources that were available at the onset of this research. It paved the way to grasp non-supervised learning basics and explore how they differ from traditional supervised learning methods. The subsequent phase involves the development of an efficient convolutional neural network (CNN) with an attention mechanism, applied to the bi-level segmentation task of road pavement crack detection. Similar to the Monkeypox classification, this is also a binary classification task, but at pixel-level, i.e., it is a two-way semantic segmentation problem. Hence, the number of samples found in the relevant datasets is once again manageable on the computational resources available during the research. [...]