Abstract: Guided image filter is a well-known local filter in image processing. However, the presence of halo artifacts is a common issue associated with this type of filter. This paper proposes an algorithm that utilizes gradient information to accurately identify the edges of an image. Furthermore, the algorithm uses weighted information to distinguish flat areas from edge areas, resulting in sharper edges and reduced blur in flat areas. This approach mitigates the excessive blurring near edges that often leads to halo artifacts. Experimental results demonstrate that the proposed algorithm significantly suppresses halo artifacts at the edges, making it highly effective for both image denoising and detail enhancement.
Introduction
Guided Image Filtering (GIF) is a fundamental technique in computer vision, widely employed for tasks such as noise reduction, HDR tone mapping, and detail enhancement. It operates by using a "guide image" to determine the smoothing behavior applied to an input image. However, standard GIF implementations frequently suffer from halo artifacts—perceptual glowing or blurred outlines that appear around high-contrast edges.
These artifacts typically arise because traditional filters apply a uniform regularization parameter ($\lambda$) across the entire image. This approach fails to differentiate between true structural edges and flat regions, leading to over-smoothing at boundaries and insufficient smoothing in uniform areas. While previous iterations like Weighted GIF (WGIF) and Gradient-Domain GIF (GDGIF) attempted to address these issues, they often struggled with fine textures or introduced secondary blurring.
Methodology
The proposed Gradient Domain Weighted Guided Image Filtering (GWGIF) algorithm improves upon the standard linear model ($Z = aG + b$) through three primary innovations designed to preserve structural integrity while minimizing noise.
1. Multi-Directional Edge-Aware Constraints
Rather than relying solely on horizontal and vertical gradients, GWGIF utilizes a global-based fast calculation for gradients in four directions. This provides a more comprehensive map of the image's texture, allowing the filter to identify fine edges that traditional methods might miss.
2. Explicit Edge Protection
The authors introduce a first-order edge protection constraint ($\eta$). In identified edge regions, this factor forces the filter coefficient ($a$) toward 1. This mathematical constraint ensures the output image preserves the guide image's structure at the edges, effectively "locking" sharpness and preventing the leakage of pixel values that causes halos.
3. Weighted Average Strategy
Standard filters often use a simple mean to handle overlapping windows, which can lead to blurring. GWGIF employs a novel strategy that calculates the absolute difference of neighborhood sums to assign weights adaptively. This allows the algorithm to distinguish flat areas from non-uniform regions during the final reconstruction phase. Despite these additions, the algorithm maintains a computational complexity of O(N), ensuring it remains as efficient as the original GIF.
Key Results and Discussion
The GWGIF algorithm was evaluated against several benchmarks, including GIF, WGIF, and GDGIF, using standard research images like "Lena" and various floral subjects.
Quantitative Performance
In denoising experiments on the "Lena" image, GWGIF achieved a PSNR of 22.65 and an SSIM of 0.9481, outperforming both the original GIF (PSNR 22.20) and WGIF (PSNR 22.43). For detail enhancement, the algorithm showed a dramatic improvement in the Blind Image Quality Index (BIQI), scoring 71.68 compared to the original GIF's 37.53.
As shown in the table above, the proposed method achieves an average gradient of 66.88, more than double that of GDGIF (27.58) and nearly five times that of WGIF (14.45). This metric directly correlates to the perceived sharpness and detail within the processed images.
Visual Quality
Qualitative analysis confirms these findings. Side-by-side comparisons in high-contrast scenarios (such as clouds and tulips) demonstrate a visible reduction in the "black edge" effect. The resulting images exhibit sharper texture definition within fine structures like petals and feathers, without the distracting glow typical of standard guided filters.
Conclusion
By integrating multi-directional gradient information with an explicit edge protection mechanism, GWGIF successfully addresses the long-standing problem of halo artifacts in guided filtering. The algorithm provides a robust balance between noise suppression and detail preservation. Future work may explore the sensitivity of the edge protection threshold across varying lighting conditions and the application of GWGIF to specialized fields such as medical imaging and satellite data processing.
Summary updated