EvTexture++: Event-Driven Texture Enhancement for Video Super-Resolution

Researchers have introduced EvTexture++, a framework that uses event-camera signals to sharpen textures in video super-resolution, marking a shift from prior work that relied on events only for motion estimation [1]. The system is described as the first event-driven architecture dedicated to texture enhancement in video super-resolution [1][2]. Event cameras capture per-pixel brightness changes at microsecond intervals, providing high-frequency spatiotemporal detail that conventional frame-based cameras miss [2]. EvTexture++ exploits that data through a customized texture enhancement branch and an iterative texture enhancement module, progressively restoring fine-grained surface detail across multiple refinement steps [1][2]. Earlier event-based super-resolution methods, such as the predecessor EvTexture, concentrated on improving flow estimation and temporal alignment [3]. EvTexture++ reorients the signal path toward texture recovery and adds a temporal texture alignment module that estimates event-guided, texture-aware flow. The module is designed to reduce texture flickering that can arise from large inter-frame motions, preserving consistency in detailed regions [1][2]. The framework is built as a plug-and-play component, meaning it can be integrated into existing video super-resolution models without retraining the entire pipeline [1][2]. In experiments across five datasets, EvTexture++ achieved state-of-the-art results [1]. When attached to recent VSR models, it delivered gains of up to 1.55 dB in peak signal-to-noise ratio on the texture-rich Vid4 dataset [1][2]. The earlier EvTexture model had reported a gain of up to 4.67 dB compared with prior event-based methods on the same dataset, underscoring the rapid progression in this subfield [3]. Code for EvTexture++ has been made publicly available on GitHub [1][2]. The work continues a broader trend in computer vision toward leveraging neuromorphic sensors for tasks that demand high temporal fidelity, though the authors note that large-scale deployment still depends on the availability of event-camera hardware and annotated multi-modal datasets [2].