Hierarchical motion magnification

Video motion magnification amplifies small, imperceptible variations, making them perceptible. It has a wide range of applications, such as measuring respiratory signals, classifying micro-expressions, measuring micro-displacements from a distance etc. Moreover, its applications are continuously expanding. Existing handcrafted techniques provide less effective solutions, often delivering lower magnification and requiring the selection of complex hyper-parameter values that vary across videos. To tackle these issues, deep-learning-based approaches are being introduced. However, they are sensitive to noise-related distortions during motion magnification. In response to these challenges, we propose a more robust hierarchical network for video motion magnification. To mitigate distortions that may arise from the magnification of noise and other illumination changes, we introduce a multi-scale manipulator with an edge-based input before magnification. Additionally, we propose a novel contrastive learning-based loss to further enhance the robustness of the magnification process against noise. To improve texture quality, we introduce a multi-resolution magnification generation architecture with a magnification decoder block. It produces hierarchical magnification from lower resolution to higher resolution. We compare the results of the proposed network both qualitatively and quantitatively with the state-of-the-art methods on synthetic and natural videos. Also, we conduct an ablation study to analyze various modules of our method. The results demonstrate that our proposed base and lightweight models outperform the current state-of-the-art methods.