Akademik Veri Yönetim Sistemi
IEEE Access, cilt.13, ss.206256-206271, 2025 (SCI-Expanded, Scopus)
Downsampling and lossy compression severely degrade image quality, making faithful detail recovery a persistent challenge for computer vision systems, especially on resource-constrained devices. We present Edge-Aware Reparameterizable Network for Compressed Image Super-Resolution (ERNCSR), a structurally reparameterizable network that couples efficiency at inference with high reconstruction fidelity. The core of ERNCSR is a lightweight multi-branch residual block that, during training, processes features through standard convolution, a fixed edge filter, and hybrid bottleneck branches to capture textures and sharp transitions. Before deployment, these branches are algebraically fused into a single 3 × 3 kernel, collapsing the model to only 13.6K parameters and enabling over 20× speedup on a consumer-grade CPU. A spatial attention module further refines features while adding negligible overhead. Training follows a two-stage curriculum: we first learn on uncompressed bicubic data, then fine-tune with dynamically simulated JPEG, WebP, and AVIF degradations using a composite loss function. Extensive evaluations on uncompressed and compressed image variants of Set5, Set14, Urban100, DIV2K, and Manga109 demonstrate consistent gains in PSNR and SSIM over comparably compact baselines, confirming ERNCSR’s suitability for real-world compressed image super-resolution.