NN In-Loop Filter Architecture Using CP-Separable Convolution

The neural network filter is divided into distinct functional blocks: head block for input processing, transition block for feature transformation, backbone blocks for hierarchical feature extraction, and tail block for output generation. Each block performs specific operations with optimized complexity, allowing the system to achieve high video coding performance while managing processing requirements through modular architecture.

Different blocks within the neural network apply specialized operations tailored to their specific functions. The head block uses specific convolution kernels for input processing, the transition block applies feature transformation operations, the backbone blocks use hierarchical feature extraction with varying complexity, and the tail block employs output generation operations. This localized optimization of operations within each block reduces overall processing complexity while maintaining high coding performance.

The neural network filter is divided into distinct functional blocks: head block for input processing, transition block for feature transformation, backbone blocks for hierarchical feature extraction, and tail block for output generation. Each block processes data in a segmented manner, reducing the peak memory requirements compared to a monolithic architecture while maintaining the comprehensive functionality needed for high video coding performance.

The architecture employs dynamic feature map dimensions and adaptive filtering operations that adjust memory usage based on the specific processing requirements of each block. The backbone blocks, in particular, use hierarchical feature extraction that dynamically manages memory by processing features at multiple scales with varying memory footprints, optimizing the balance between performance and memory consumption.