Loop Restoration Stage in AV1 Decoding Explained

The Loop Restoration stage in the AV1 video codec is an advanced in-loop filtering tool designed to reverse the blur, ringing, and loss of fidelity introduced during block-based compression and earlier filtering stages. By applying mathematically driven restoration algorithms to reconstructed frames, this stage recovers high-frequency details and edge sharpness, significantly improving the visual quality of the output and providing cleaner reference frames for future inter-frame prediction.

Position Within the AV1 Decoding Pipeline

AV1 utilizes a multi-stage in-loop filtering architecture to reconstruct video frames before they are stored in the decoded picture buffer. The decoding process applies filters in the following strict order:

1. Inverse Quantization and Transform: Reconstructs the raw prediction and residual data.
2. Deblocking Filter: Smooths grid-like discontinuities across transform block boundaries.
3. Constrained Directional Enhancement Filter (CDEF): Identifies directional edges to suppress ringing artifacts while preserving structural lines.
4. Loop Restoration: Applies mathematical deblurring to restore fine textures and edges lost in previous steps.
5. Film Grain Synthesis: An optional post-processing step applied outside the reference frame loop to reintroduce natural film texture.

Because Loop Restoration operates as the final in-loop filter, its output directly forms the reference frames that subsequent predictive frames rely on.

Primary Functions of Loop Restoration

While Deblocking and CDEF remove unwanted high-frequency artifacts (such as block borders and ringing noise), they also unavoidably blur authentic textures and soften sharp edges. The Loop Restoration filter counters this degradation by acting as an inverse filter.

During the encoding process, the encoder compares the degraded, pre-filtered frame to the pristine original source frame. It calculates correction parameters (filter coefficients) that bring the reconstructed frame mathematically closer to the original. These parameters are signaled in the bitstream, allowing the AV1 decoder to apply precise restorations with minimal computational overhead.

Restoration Methods in AV1

The AV1 specification defines two primary filter algorithms for the Loop Restoration stage, alongside a bypass option:

• Separable Symmetric Wiener Filter: This is a 2D separable linear filter (using horizontal and vertical taps) that mathematically approximates the inverse of blur. The encoder calculates Wiener coefficients that minimize the mean squared error between the original image and the reconstructed image. The decoder applies these coefficients to restore lost sharpness.
• Dual Self-Guided Filter: Based on domain transform and guided image filtering techniques, this algorithm performs edge-preserving smoothing. It uses two distinct radius and noise parameters to create guided projections of the image, combining them to boost high-frequency features without creating overshooting or "halo" artifacts around sharp edges.
• None (Identity): The encoder can disable restoration for specific areas where processing would not yield noticeable visual or compression gains, conserving computational resources.

Restoration Units

Loop Restoration does not operate strictly on transform or coding blocks. Instead, it processes the image in larger, configurable regions called Restoration Units (RUs).

Restoration Units are typically sized at 64x64, 128x128, or 256x256 pixels for the luma plane, and can be configured independently for chroma planes. Each Restoration Unit can independently select between the Wiener filter, the Dual Self-Guided filter, or no filter at all. This unit-based structure provides a high degree of spatial adaptivity, allowing the decoder to apply intense sharpening to high-detail areas while leaving flat or smooth regions untouched.

Impact on Video Quality and Compression

The Loop Restoration stage delivers substantial coding efficiency gains, often accounting for a significant portion of AV1’s overall bitrate reduction compared to older codecs like VP9 and HEVC. By creating higher-fidelity reference frames, subsequent motion compensation requires smaller residual errors, allowing the encoder to compress subsequent frames more aggressively without degrading final picture quality.