Improving Image Compression Ratio Using the mUMamba Image Restoration Network

Abstract The exponential growth of multimedia content sharing, with approximately 3 billion images transmitted daily across digital platforms, necessitates efficient compression techniques that balance storage reduction with reconstruction quality. This paper addresses the fundamental challenge of achieving superior compression ratios while maintaining perceptual fidelity through mUMamba, a novel image compression framework that synergistically integrates hierarchical block removal strategies with Mamba-based neural reconstruction. The proposed methodology employs a multi-scale adaptive analysis paradigm that progressively evaluates image blocks at three granularities (16×16, 8×8, and 4×4 pixels) to identify and remove redundant regions based on edge complexity and variance characteristics. Removable blocks are classified into edge-containing regions, where only edge patterns are transmitted, and gradient-characterized regions, where directional gradient vectors serve as auxiliary reconstruction information. The decoder implements a modified U-Net architecture incorporating Vision Mamba blocks within deeper layers, achieving efficient long-range dependency modeling through parallel state-space operations, while compression-aware Channel Attention and Spatial Attention Bridges enhance skip connections for superior feature fusion. Comprehensive experimental validation on multiple benchmark datasets (CLIC, Kodak, Tecnick, DIV2K) demonstrates that mUMamba achieves a compression ratio of 46.5% at a bitrate of 0.75 bits per pixel (bpp) while maintaining a PSNR of 25.15 dB and an SSIM of 0.791, representing a BD-rate improvement of -18.2% over BPG/VVC standards. Comparative analysis reveals 22.8% SSIM improvement and 7.0% PSNR gain over state-of-the-art autoregressive methods (BINetAR, ConvAR), while simultaneously reducing model parameters by 80% (8.5M vs. 42.5M), computational complexity by 38.8% (32.0G vs. 52.3G FLOPs), and inference time by 31.6% (65 ms vs. 95 ms). Ablation studies confirm that the hierarchical multi-scale strategy yields 32% higher compression ratio compared to single-scale baselines with counterintuitive quality improvements through adaptive preservation of complex regions. These results establish mUMamba as a computationally efficient solution for resource-constrained deployment scenarios including mobile platforms, IoT sensor networks, telemedicine systems, and content delivery networks where both compression efficiency and reconstruction fidelity are critical requirements.