摘要
This paper presents CATTUS, the first deep image processor to integrate U-Net acceleration with channel attention, enabling high-quality image deblurring on mobile devices. It achieves throughput of 4K-UHD 30fps with 4.4 TOPS/W of energy efficiency. The processor realizes image reconstruction on challenging tasks such as deblurring with three key features: 1) pyramid layer fusion (PLF) for efficient U-net acceleration, 2) skip tunneling that minimizes external memory access (EMA) by optimizing skip connections, 3) lookahead channel attention (LCA) to reduce EMA for aggregating global information. Fabricated in 16nm FinFET technology, CATTUS delivers a 1.53× throughput improvement and 82% EMA reduction compared to baseline implementation, demonstrating real-time imaging across multiple applications.