Self-Supervised Learning for Image Classification using Contrastive Consistency Learning

Self-Supervised Learning for Image Classification is an advanced AI-powered computer vision project that reduces dependency on large labeled datasets by learning meaningful image representations from unlabeled data. The system implements a two-stage learning pipeline consisting of self-supervised pre-training followed by supervised fine-tuning using the CIFAR-10 dataset. The project utilizes a ResNet-18 encoder with a contrastive consistency learning approach, enabling the model to learn robust feature representations from unlabeled images through consistency loss. During fine-tuning, only a small portion of labeled data is required, significantly reducing annotation costs while maintaining competitive classification performance. A secure Flask-based web application allows authenticated users to upload images and receive real-time predictions across the ten CIFAR-10 classes. The application includes user registration, login, session management, password hashing, image preprocessing, and prediction visualization. The project demonstrates modern deep learning concepts including transfer learning, self-supervised representation learning, feature extraction, model deployment, secure authentication, and web-based AI inference. It serves as an excellent IEEE/final-year project for students interested in Artificial Intelligence, Machine Learning, Deep Learning, and Computer Vision.