GSENFlow: Learning Optical Flow with Globally Separable Enhanced Noise

Optical flow estimation, a core task in the field of computer vision, provides valuable motion information by tracking the movement of pixels between adjacent frames in video or image sequences. This information is not only crucial for revealing dynamic changes in scenes but also forms the foundation for advanced applications such as video super-resolution, video frame interpolation, autonomous driving, and robotic navigation. Despite the self-evident importance of optical flow, maintaining the accuracy of models in complex scenes, especially under motion blur conditions, has always been a significant challenge. To address this challenge, this chapter proposes an innovative robust optical flow algorithm based on video spatiotemporal features called GSENFlow. The algorithm draws on the GMFlow framework and integrates Enhanced Multi-Scale Features (EMA) modules and Deep Separable Convolutional Networks (SeNet) modules to enhance the network’s ability to capture multi-scale features in complex scenes. Moreover, we introduce a noise layer to simulate real conditions under motion blur, thereby improving the algorithm’s robustness. Experimental results demonstrate that our proposed GSENFlow algorithm exhibits outstanding performance in handling scenes with rapid motion and motion blur. These achievements not only advance the research progress of optical flow algorithms but also provide strong technical support for practical applications in related fields, broadening the application prospects of optical flow technology.