Document Type

Article

Publication Date

5-22-2026

First Advisor

Kyle Wilson

Abstract

Single Image Super-Resolution (SISR) reconstructs high-resolution images from low-resolution inputs, a capability valuable across many application domains. This thesis focuses on the replication and architectural enhancement of the Omni Aggregation Network (OAN), a lightweight super-resolution architecture designed to balance reconstruction quality with computational efficiency through a novel omni-axis self-attention mechanism.

Independent replication remains a persistent challenge in machine learning research, and this work addresses that issue directly. Initial replication attempts trailed the original reported results by approximately 6-10% in performance metrics before several undocumented implementation details were identified. After correcting these discrepancies, the replicated implementation achieved results within 4% of the reported metrics across standard benchmark datasets.

Building on this validated baseline, three architectural modifications are proposed and evaluated: a perceptual and adversarial fine-tuning stage, replacement of Enhanced Spatial Attention with Coordinate Attention, and a reduced-width, increased-depth restructuring strategy. Experimental results show that Coordinate Attention provides the most effective improvement, increasing both PSNR and SSIM with only minimal latency overhead.

Download

Share

COinS