💻 computer science

NPDO: Neural Prediction Direction Optimizer for Fast VVC Intra Coding

This paper proposes NPDO, a hybrid neural framework combining multi-scale feature extraction with hierarchical CNN and Vision Transformer architectures to intelligently prune the VVC intra-prediction search space, achieving a 54.0% reduction in encoding time with only a 1.18% BD-Rate increase compared to the VTM 23.0 reference.

Original authors: Reka Sandaruwan Gallena Watthage, Anil Fernando

Published 2026-07-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Reka Sandaruwan Gallena Watthage, Anil Fernando

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you're trying to pack a massive, chaotic attic into a tiny suitcase for a move. The attic represents a high-definition video frame, and the suitcase is the compressed file you want to send over the internet. The challenge? The attic is full of 67 different ways to fold and stack the items (called "prediction modes"). To find the perfect way to pack everything so it takes up the least space without breaking anything, a standard video coder (like the VTM 23.0 reference) tries every single one of those 67 ways. It's like trying every possible combination of folding a shirt, every single time you pack a box. It works, but it's incredibly slow and burns a lot of energy.

Enter NPDO (Neural Prediction Direction Optimizer), a new "smart assistant" designed to speed up this packing process for the latest video standard, VVC.

The Problem with the Old Way
The paper argues that the traditional method of checking all 67 folding directions is too slow for real-time use, especially for 4K video. It's like a librarian who insists on reading every single book in the library to find the one you need, rather than just checking the most likely shelves. The authors explicitly rule out relying solely on old-school "hand-crafted" rules (like simple math tricks to guess the direction) or using just one type of computer brain (like only a standard camera-like neural network). They found these older methods either miss the big picture or get confused by complex textures.

The NPDO Solution: A Two-Brain Detective
Instead of checking everything, NPDO acts like a super-smart detective that looks at the "attic" and instantly guesses the top few folding methods that will work best. It does this using a clever three-step process:

The Multi-Scale Scan: First, it uses a "magic lens" called a Discrete Wavelet Transform (DWT) to see the texture of the video at different sizes, like looking at a forest from a drone and then from the ground. It also draws a map of the edges (like the outlines of trees) using a histogram.
The Two-Brain Team: This information is fed into two different types of artificial intelligence working together.
- The Local Expert (HCNN): A Hierarchical Convolutional Neural Network that is great at spotting small, local patterns, like noticing a specific leaf shape.
- The Global Expert (ViT): A lightweight Vision Transformer that looks at the whole picture to understand the big picture, like seeing the entire forest layout.
- These two brains talk to each other and combine their opinions to make a final decision.
The Smart Filter: Finally, the system looks at how "messy" or complex the video area is. If the area is simple (like a blue sky), it only checks 3 folding options. If it's complex (like a busy crowd), it checks 5. It never wastes time checking all 67.

What the Numbers Say
The authors ran extensive tests on thousands of video sequences to see if this actually works. They didn't just guess; they measured the results against the standard VTM 23.0 encoder.

Speed: NPDO cuts the time it takes to encode video by 54.0%. That means it's more than twice as fast.
Quality: The trade-off is tiny. The video quality drops by only 1.18% in terms of bitrate efficiency (a metric called BD-Rate).
Accuracy: The system is incredibly good at guessing. It picks the correct top 5 folding directions 98.1% of the time.
Efficiency: Instead of testing 67 options, it only tests an average of 4.2 options per block, reducing the work by 93.7%.

What It's Not
The paper is careful to note that while NPDO is fast, it doesn't replace the entire video coding system; it just speeds up the "intra prediction" part (predicting how a block looks based on its neighbors). Also, while it works great on 4K video, the authors note that lower-resolution videos (like Class D) see slightly less speedup because there is less texture information to analyze, though it still beats the old methods.

The Bottom Line
In these simulations and tests, NPDO proves that you don't need to check every single door to find the treasure. By using a hybrid team of AI brains that look at both small details and the big picture, it can skip the vast majority of useless guesses. The result? You can encode 4K video in real-time (30 frames per second) without the computer getting exhausted, all while keeping the picture quality almost identical to the slow, exhaustive method. It's a significant step forward, turning a slow, manual packing job into a lightning-fast, automated one.

Technical Summary: NPDO: Neural Prediction Direction Optimizer for Fast VVC Intra Coding

Problem Statement

Versatile Video Coding (VVC/H.266) achieves approximately 50% bitrate reduction compared to its predecessor, HEVC, but at the cost of significantly increased computational complexity. A primary bottleneck is the intra-prediction framework, which supports up to 67 prediction modes (including Planar, DC, 65 angular modes, and cross-component linear models). The standard encoder must exhaustively evaluate all modes via Rate-Distortion Optimization (RDO) for every Coding Unit (CU), a process that typically requires 10–20 times more computational resources than HEVC. This exhaustive search severely hinders real-time encoding implementations, particularly for high-resolution content like 4K.

Existing fast encoding methods face several limitations:

Fragmented Feature Representation: Traditional approaches often rely solely on frequency-domain analysis (e.g., DWT) or spatial gradient features, failing to integrate both within a unified framework.
Static Decision Mechanisms: Many methods use fixed thresholds regardless of content complexity, leading to over-computation on simple textures or quality degradation on complex regions.
Architectural Limitations: Current deep learning solutions typically rely exclusively on either Convolutional Neural Networks (CNNs) for local patterns or Transformers for global dependencies, missing the opportunity to leverage their complementary strengths.
Integration Challenges: Few frameworks offer seamless scalability or maintain full standard compatibility without requiring substantial encoder modifications.

Methodology

The paper proposes the Neural Prediction Direction Optimizer (NPDO), a hybrid framework designed to intelligently prune the prediction search space before the RDO stage. The system operates in five sequential stages:

1. Feature Extraction Engine

NPDO constructs a unified 30-dimensional feature vector by synergistically combining three complementary modules:

Discrete Wavelet Transform (DWT) Module: Decomposes the CU using 4-level Haar or Daubechies-4 wavelets. It calculates energy values for four subbands (LL, LH, HL, HH) to capture multi-scale texture orientation.
Histogram Module: Applies Canny edge detection and Sobel operators to compute gradient magnitudes and orientations. These are accumulated into a 16-bin histogram (22.5° per bin) to capture local directional edge information.
Context Module: Exploits spatial correlation by incorporating prediction modes from neighboring CUs (Left, Top, Top-Left) to construct the Most Probable Mode (MPM) list. It also includes geometric features (CU size, depth, aspect ratio) and texture complexity metrics (variance, entropy).

2. Dual Neural Network Architecture

The extracted features are processed by two parallel neural architectures to generate confidence scores for all 67 intra modes:

Hierarchical CNN (HCNN): A 3-stage residual architecture optimized for efficient local pattern recognition. It expands channel dimensions (64 → 128 → 256) using 1D convolutions, batch normalization, and ReLU activations, concluding with global average pooling and a fully connected layer.
Lightweight Vision Transformer (ViT): Designed to capture long-range dependencies. It adapts patch sizes based on CU dimensions (8×8 for larger CUs, 4×4 for smaller ones), utilizes multi-head self-attention (4 heads), and processes a classification token ([CLS]) to aggregate global context.

3. Adaptive Fusion and Directional Attention Index (DAI)

The outputs of the HCNN and ViT are fused using a learned adaptive weighting parameter ( $\alpha = 0.6$ ) to produce a final confidence score for each mode. This generates the Directional Attention Index (DAI), a set of mode-confidence pairs.

4. Content-Adaptive Top-K Selection

Instead of a fixed pruning strategy, NPDO calculates a Complexity Score based on variance, gradient magnitude, DCT energy, and neighbor consistency. Based on this score, the system dynamically adjusts the number of candidate modes ( $K$ ) and a confidence threshold ( $\tau$ ):

Simple Content ( $S < 0.3$ ): $K=3$ , $\tau=0.7$
Medium Content ( $0.3 \le S < 0.6$ ): $K=4$ , $\tau=0.6$
Complex Content ( $S \ge 0.6$ ): $K=5$ , $\tau=0.5$

5. Selective RDO

The encoder evaluates only the top- $K$ candidates identified by the DAI through the standard RDO process, rather than all 67 modes.

Key Contributions

The paper identifies four primary contributions:

Hybrid Feature Extraction: A comprehensive engine combining DWT-based multi-scale analysis, gradient-oriented histograms, and spatial context features to effectively capture directional content characteristics.
Complementary Neural Architectures: The integration of a Hierarchical CNN for local patterns and a lightweight Vision Transformer for global dependencies, unified through an adaptive learned fusion mechanism.
Directional Attention Index (DAI) with Adaptive Selection: A novel scheme that dynamically adjusts the candidate mode set size based on texture complexity, achieving high pruning efficiency (92.5–95.5%) while maintaining 98.1% top-5 prediction accuracy.
Experimental Validation: Extensive testing demonstrating significant encoding time reductions with minimal quality degradation, enabling real-time 4K VVC encoding.

Experimental Results

The framework was evaluated using the JVET Common Test Conditions and other diverse datasets (UVG, BVI-DVC) against the VTM 23.0 reference encoder and state-of-the-art baselines.

Encoding Time Reduction: NPDO achieved an average encoding time reduction of 54.0% compared to VTM 23.0.
Rate-Distortion Performance: The method incurred a minimal 1.18% increase in BD-Rate.
Comparison to Baselines:
- Compared to a DWT/Histogram baseline (50.5% time reduction, 1.53% BD-Rate increase), NPDO provided a 3.5% improvement in encoding speed and superior rate-distortion performance.
- Compared to machine learning approaches like Tissier et al. [17] (43.5% time reduction, 2.26% BD-Rate increase), NPDO demonstrated consistent improvements across all video classes.
Prediction Accuracy: The framework achieved 98.1% top-5 prediction accuracy, ensuring the optimal mode is included in the candidate set for the vast majority of CUs.
Complexity Reduction: Mode evaluations per CU were reduced by 93.7% (from 67 to an average of 4.2). This resulted in a 2.17× speedup, a 40% reduction in memory usage, and a 34% reduction in power consumption.
Real-Time Capability: The system enables real-time 4K encoding at 30fps with 46% CPU utilization.

Significance and Claims

The paper claims that NPDO represents a significant advancement toward practical deployment scenarios, including cloud encoding, mobile applications, and broadcast systems. By successfully integrating traditional signal processing (DWT, histograms) with learned representations (HCNN, ViT), the framework addresses the "fragmented feature representation" gap in existing literature.

The authors emphasize that the content-adaptive nature of the top-K selection is crucial; it allows for aggressive pruning on simple textures without sacrificing accuracy on complex regions, a balance that static threshold methods fail to achieve. The results validate that combining local and global neural architectures yields a superior complexity-quality trade-off compared to using either architecture alone. The framework maintains full compatibility with the VVC standard, requiring no modifications to the core coding tools other than the pre-RDO mode selection logic, facilitating widespread adoption.