Abstract: This research presents a refined approach to image super-resolution (SR) by integrating edge-preserving initial estimates with learning-based residual compensation. By utilizing Histogram of Shearlet Coefficients (HSC) descriptors and Locally-Linear Embedding (LLE), the proposed method significantly reduces visual artifacts and improves reconstruction accuracy compared to traditional state-of-the-art techniques.
Introduction
High-resolution (HR) images are indispensable in specialized fields such as medical imaging, satellite surveillance, and industrial automation. However, acquiring such images is often hindered by the high cost of precision sensors and physical constraints on pixel size reduction, which can introduce significant noise. Super-resolution (SR) algorithms address these challenges by computationally estimating HR images from low-resolution (LR) inputs.
Existing SR methods frequently struggle with two primary issues: blurring and the introduction of unnatural artifacts. Simple interpolation techniques, like bicubic upscaling, tend to lose sharp edge information. Conversely, learning-based methods that attempt to synthesize missing details can produce "ghosting" or blocky textures, particularly when relying on raw pixel values for patch matching in training databases.
Proposed Methodology
The core of the proposed method is an iterative pipeline focused on residual compensation. Instead of directly estimating the HR image, the algorithm learns the difference (residual) between a blurry upscaled image and its true HR counterpart.
The process begins with an initial estimation using Gradient Profile Prior (GPP). Unlike standard interpolation, GPP focuses on preserving the sharpness and shape of gradient profiles (edges). Following this, the method performs feature extraction using Histogram of Shearlet Coefficients (HSC). Shearlets are particularly effective at capturing multi-scale, directional information, making them more robust than raw pixels for representing image structure.
The learning phase utilizes Locally-Linear Embedding (LLE) to identify the $k$-nearest neighbors of the LR residual's HSC descriptors within a pre-trained database. Optimal weights are calculated to reconstruct the corresponding HR residual, which is then added back to the image. This refinement process iterates until the Root Mean Square Error (RMSE) falls below a target threshold or the maximum number of iterations is reached.
Experimental Results
The method was evaluated using the FERET face database, a standard benchmark for facial image processing. The training set comprised 228 HR images, while 20 images served as the test set. LR versions were created by downsampling images to $84 \times 84$ pixels after eye alignment.
Quantitative analysis for $2\times$ magnification showed that the proposed method achieved an average RMSE of 9.8276, outperforming both GPP (10.0383) and previous LLE-based methods (10.8438). Qualitative results further demonstrate the method's superiority in preserving fine details.
As shown in the visual comparisons, the proposed approach produces smoother textures and sharper features. Notably, it effectively eliminates blocky or noisy artifacts around sensitive facial regions like the eyes and mouth, which are common failure points for other learning-based algorithms. Parameter tuning revealed that using $k=5$ neighbors and a smoothing factor ($\alpha$) of 0.01 yielded the most stable and accurate results.
Conclusion and Future Work
The integration of HSC descriptors into a residual-learning framework provides a robust solution for high-quality image super-resolution. By focusing on structural features rather than raw intensities, the method achieves significant reductions in both reconstruction error and visual artifacts.
Future research will explore the generalizability of these descriptors to other domains, such as medical or satellite imagery, and investigate optimizations to reduce the computational overhead associated with the iterative LLE search process.
Summary updated