Contrastive Self-Supervised Vision Model for Minimizing Image Recognition Annotation Necessities

Abstract

Deep learning techniques (which enable systems to recognize automatically the complex visual patterns have significantly improved image recognition. However, these models typically have the large labelled data sets (which are expensive), and time-consuming and requiring special knowledge to produce, especially in manufacturing, healthcare and environmental monitoring. Self-Supervised Learning (SSL) overcomes this problem by eliminating manual labeling by use of pretext tasks aimed at acquiring relevant features to directly generate supervisory information from data. In order to train an encoder on unlabeled images, we take a look at SimCLR, a contrastive SSL framework to maximizes similarity between two augmented views of same image discriminates it from others. A simple linear classifier on a small fraction of labeled set is then, we used for evaluating the pretrained encoder. This technique shows that SSL can be made to have competitive accuracy and significantly reducing the annotated data dependency and makes it a viable and practical choice for real-world applications where labeled data is scarce. Experimental results show that the decrease in the demand of annotation has a direct impact on training efficacy and generalization functioning.