Adaptive Loop Filter Virtual Boundary Processing for Lower Signaling Overhead

The patent segments pixels into 25 distinct classes based on local neighborhood characteristics (gradients and activity). Instead of processing every pixel individually, the encoder classifies each pixel into one of 25 groups and signals filter coefficients at the class level. This segmentation reduces the number of coefficients that need to be transmitted while maintaining adaptive filtering capability across different local regions.

The patent changes the parameter representation from per-pixel coefficients to per-class coefficients. By transforming the filtering parameters from a fine-grained (per-pixel) to a coarser-grained (per-class) representation, the system reduces transmission overhead. The class assignment serves as an additional parameter that enables efficient coefficient reuse while preserving the ability to adapt to local image characteristics.

The patent applies local quality by creating 25 distinct pixel classes based on local neighborhood characteristics such as gradients and activity. Each class represents a specific local pattern, allowing filter coefficients to be adapted to local image features. The classification process ensures that pixels with similar local characteristics share coefficients, while pixels with different characteristics have different coefficients, thereby maintaining local adaptability.

The patent introduces dynamic adaptability through the classification mechanism. The system dynamically determines which of the 25 classes each pixel belongs to based on local image characteristics, and consequently dynamically selects appropriate filter coefficients. This dynamic class-based selection provides adaptability comparable to per-pixel filtering while reducing transmission overhead, as the adaptability is achieved through intelligent grouping rather than individual processing.