This technology is a method for executing edge-adaptive intra-field deinterlacing in parallel on a graphics processing unit (GPU). It estimates a deinterlacing prediction model using six neighboring pixels around the target pixel based on geometric duality. By assigning one thread to each missing pixel, it interpolates a vast number of pixels simultaneously.
Interlaced broadcast formats like NTSC and PAL often suffer from flickering, artifacts, and field aliasing, while simple line-average interpolation tends to blur vertical edges. Although covariance-based adaptive deinterlacing preserves edge sharpness and provides superior image quality, its high computational demand—requiring a prediction model to be solved for every pixel—has historically limited real-time processing.
This technology distributes interlaced images and intermediate variables across multiple banks of GPU shared memory, allowing simultaneous thread access through block-level synchronization, while storing auto-generated variables in register memory to reduce access latency. It further minimizes transfer time by overlapping kernel operations and host transfers using asynchronous data streams and multi-GPU utilization. Applicable to digital TV receivers and broadcast video converters, it enables high-quality adaptive deinterlacing at real-time broadcast speeds.
N/A
