Hitchlearning: A general free-lunch paradigm for single-image enhancement by unifying inference and training
Welcome to the official GitHub repository of HitchLearning, an innovative Deep Learning (DL) paradigm that revolutionizes the traditional approach by merging training and inference into a cohesive process.
📰 Paper 🏗️ Model (via Google) 🏗️ Model (via Baidu) 🗃️ Dataset 🧱 Code 🧐 Video 🧑💻 Demo
HitchLearning is not just another DL methodology. It's a groundbreaking approach that challenges the conventional separation of training and inference found in traditional DL paradigms. The core innovation of HitchLearning lies in its rejection of the assumption that training and inference data are independently and identically distributed (i.i.d.).
- Unified Training and Inference: HitchLearning uniquely optimizes the model using a single inference image, thereby unifying the process of training and inference. This method addresses the non-i.i.d problem inherent in many DL applications.
- Real-Time Adaptation: By focusing on each specific inference image, HitchLearning adapts in real-time to its unique characteristics, offering a significant improvement in model performance.
- Efficient and Effective: This paradigm improves model performance in a 'free-lunch' manner, as it requires no additional data other than the single inference image for optimization.
- Extensive Evaluation: HitchLearning has been rigorously tested across three different tasks - denoising, deblurring, and super-resolution. It has shown remarkable results in both supervised and unsupervised models across four diverse datasets.