[Paper Review] Regression-based Intra-prediction for Image and Video Coding

Carlo Noel Ochotorena, Yukihiko Yamashita|arXiv (Cornell University)|May 12, 2016

Video Coding and Compression Technologies23 references3 citations

TL;DR

This paper proposes a regression-based intra-prediction (RIP) method that iteratively refines HEVC-style prediction modes using regularized regression on natural image blocks, significantly improving prediction accuracy across all block sizes and conditions—especially in low-bitrate, worst-case scenarios—while maintaining low computational complexity during encoding via matrix multiplication.

ABSTRACT

By utilizing previously known areas in an image, intra-prediction techniques can find a good estimate of the current block. This allows the encoder to store only the error between the original block and the generated estimate, thus leading to an improvement in coding efficiency. Standards such as AVC and HEVC describe expert-designed prediction modes operating in certain angular orientations alongside separate DC and planar prediction modes. Being designed predictors, while these techniques have been demonstrated to perform well in image and video coding applications, they do not necessarily fully utilize natural image structures. In this paper, we describe a novel system for developing predictors derived from natural image blocks. The proposed algorithm is seeded with designed predictors (e.g. HEVC-style prediction) and allowed to iteratively refine these predictors through regularized regression. The resulting prediction models show significant improvements in estimation quality over their designed counterparts across all conditions while maintaining reasonable computational complexity. We also demonstrate how the proposed algorithm handles the worst-case scenario of intra-prediction with no error reporting.

Motivation & Objective

• Address the limitation of handcrafted intra-prediction modes in HEVC and AVC, which do not fully exploit natural image structures.
• Improve prediction accuracy by learning linear mappings from reference pixels to predicted blocks using real image data.
• Develop a robust intra-prediction system that performs well even under worst-case conditions, such as no error reporting and only top-left block encoded.
• Maintain low computational complexity during encoding by reducing intra-prediction to a single matrix multiplication after offline training.
• Demonstrate the extensibility of the method to video coding by isolating intra-prediction performance in image coding.

Proposed method

• Seed the algorithm with existing HEVC-style prediction modes (DC, planar, and 33 angular modes) as initial predictors.
• Use an iterative refinement process inspired by k-means clustering: assign each training block to the best-performing predictor, then retrain a new linear regression model for each cluster.
• Apply regularized regression (e.g., ridge regression) to prevent overfitting and improve generalization across image structures.
• Construct a global prediction matrix by stacking group-specific regression matrices, enabling single-step prediction via matrix multiplication.
• Leverage optimized linear algebra libraries (e.g., BLAS) to ensure efficient inference during encoding.
• Enable worst-case intra-coding by reconstructing the entire image using only the top-left block and RIP estimates, with residual errors reported only once.

Experimental results

Research questions

• RQ1 Can a data-driven regression approach significantly improve intra-prediction accuracy compared to handcrafted HEVC-style modes?
• RQ2 How does the proposed method perform under worst-case intra-coding conditions with no error reporting and minimal prior information?
• RQ3 To what extent can iterative refinement of regression models enhance prediction quality across diverse image content and block sizes?
• RQ4 Does the RIP method maintain low computational complexity during encoding despite its improved accuracy?
• RQ5 Can the learned predictors generalize effectively across different image types and block dimensions (e.g., 4×4 to 32×32)?

Key findings

• The RIP method achieves an average improvement of 5.747 dB, 6.693 dB, and 2.432 dB in prediction quality over conventional HEVC modes for the Lena, Peppers, and Mandrill images, respectively, under best-case conditions.
• For 32×32 blocks, RIP maintains a strong quantitative lead over HEVC-style predictors, though subjective evaluation shows increased noise due to the large dimensionality gap between reference and predicted pixels.
• In worst-case intra-coding scenarios—where only the top-left block is encoded and all others are predicted without error reporting—the RIP method produces a stable, low-complexity reconstruction that enables efficient residual encoding.
• The method reduces the need for rate-distortion optimization during intra-coding by enabling a serial, error-reporting-free reconstruction pipeline.
• The final prediction step is computationally efficient, requiring only a single matrix multiplication per block, thanks to pre-trained regression matrices.
• The approach is extensible to video coding, as the core intra-prediction mechanism is independent of temporal prediction and relies solely on spatial correlation.