RCENet: Recursive Concatenation and Enhancement Network for Real-Time Super-Resolution

Recent advancements in edge AI have increased demand for real-time vision models that run efficiently on edge devices. However, their architectural heterogeneity in terms of compute structure, memory bandwidth, and supported tasks requires a vision model optimized for each edge device. Therefore, we present the Recursive Concatenation and Enhancement Network (RCENet), a lightweight and efficient Single Image Super-Resolution (SISR) model optimized for Google Tensor Processing Units (TPUs). To optimize the architecture for Google TPUs, we first conduct a detailed analysis of computational characteristics and runtime behavior to inform the network design. As a result, RCENet leverages hardware-efficient operators and quantization-friendly modules. We further propose Operator-Selective Quantization (OSQ) combined with Quantization-Aware Distillation (QAD), tailored to the TPU architecture, to enable deployment on integer-only inference engines without compromising perceptual quality. Extensive experiments on standard benchmarks show that RCENet delivers competitive visual quality with significantly reduced latency and power consumption. In particular, RCENet achieves more than 70 FPS on Google TPUs, while maintaining visual quality comparable to that of much more complex models. Our method achieved second place on the Quantized Super-Resolution track of the 2025 Mo-bileAI (MAl) Challenge, demonstrating its effectiveness for real-world deployment. Our project page is available at https://rlghksdbs.github.ioIRCENet/