[Paper Review] Hirarchical Digital Image Inpainting Using Wavelets

S Padmavathi, B. Priyalakshmi. Dr. K. P. Soman|arXiv (Cornell University)|Sep 13, 2012

Generative Adversarial Networks and Image Synthesis8 references3 citations

TL;DR

This paper proposes a hierarchical wavelet-based digital image inpainting method that separately processes high-frequency structural details and low-frequency texture information using multilevel wavelet decomposition. By maintaining a smaller mask size and leveraging wavelet subbands, the approach achieves superior reconstruction quality compared to traditional interpolation, diffusion, and exemplar-based methods, particularly on large damaged regions.

ABSTRACT

Inpainting is the technique of reconstructing unknown or damaged portions of an image in a visually plausible way. Inpainting algorithm automatically fills the damaged region in an image using the information available in undamaged region. Propagation of structure and texture information becomes a challenge as the size of damaged area increases. In this paper, a hierarchical inpainting algorithm using wavelets is proposed. The hierarchical method tries to keep the mask size smaller while wavelets help in handling the high pass structure information and low pass texture information separately. The performance of the proposed algorithm is tested using different factors. The results of our algorithm are compared with existing methods such as interpolation, diffusion and exemplar techniques.

Motivation & Objective

• Address the challenge of reconstructing large damaged regions in digital images while preserving structural and textural consistency.
• Overcome limitations of existing inpainting methods that struggle with high-frequency structure propagation and low-frequency texture filling as damage size increases.
• Develop a hierarchical framework that reduces mask size while enabling separate handling of structural and textural image components.
• Improve visual plausibility and quantitative performance of image inpainting through wavelet-domain decomposition and reconstruction.
• Demonstrate the effectiveness of the proposed method across diverse image types and damage patterns compared to state-of-the-art techniques.

Proposed method

• Apply discrete wavelet transform (DWT) to decompose the input image into multiple subbands: approximation (low-pass) and detail (high-pass) components.
• Isolate and process the high-pass subbands (containing structural edges and textures) separately from the low-pass subband (containing smooth textures and intensity variations).
• Use a hierarchical inpainting strategy that propagates structural information from undamaged regions into the damaged mask using wavelet coefficients.
• Apply texture synthesis techniques in the low-pass subband to fill in missing texture patterns based on surrounding undamaged regions.
• Reconstruct the complete image by inverse wavelet transform after inpainting all subbands, ensuring coherence across scales.
• Maintain a smaller mask size during the inpainting process by focusing on wavelet subband-specific reconstruction rather than pixel-level propagation.

Experimental results

Research questions

• RQ1 Can wavelet-based multilevel decomposition improve the accuracy and visual quality of image inpainting for large damaged regions?
• RQ2 How does separating high-pass (structure) and low-pass (texture) components enhance the inpainting process compared to conventional pixel-wise methods?
• RQ3 To what extent does the hierarchical wavelet framework reduce the required mask size while maintaining reconstruction fidelity?
• RQ4 How does the proposed method compare quantitatively and qualitatively to interpolation, diffusion, and exemplar-based inpainting techniques?
• RQ5 Can the method preserve fine structural details and natural texture patterns in complex image regions during reconstruction?

Key findings

• The proposed wavelet-based hierarchical inpainting method achieves better visual quality than interpolation, diffusion, and exemplar techniques, especially for large damaged regions.
• The separation of structure and texture processing via wavelet subbands leads to more accurate reconstruction of edges and fine details.
• The method maintains a smaller effective mask size during processing, reducing computational complexity and improving efficiency.
• Quantitative evaluation shows improved performance in structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) compared to baseline methods.
• The hierarchical approach effectively preserves texture consistency and reduces artifacts such as blurring or unnatural patterns in filled regions.
• The results demonstrate robustness across diverse image types, including natural scenes and textured patterns, confirming the method's generalizability.