Recent advancements in edge AI have increased demand for real-time vision models that run efficiently on the edge devices. However, the architectural heterogeneity of edge devices in terms of compute structure, memory bandwidth, and supported tasks requires a computer vision network architecture that is optimized for the individual 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 the 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 Tensor 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. Notably, RCENet achieves over 70 FPS on Google Tensor NPUs while matching the quality of much heavier models. Our method achieved second place in the Quantized Super-Resolution track of the 2025 MobileAI (MAI) Challenge, demonstrating its effectiveness for real-world deployment.
RCENet Overall architecture
This work was supported by The KEIT(Korea planning \& Evaluation Institute of Industrial Technology) grant funded by the Korea government(MOTIE) (No.20014107, Development of 3D camera technology with fusion of thermal images to cope with the night and day environment)