MobiSR: Efficient On-Device Super-Resolution through Heterogeneous Mobile Processors
This work addresses the problem of on-device image upscaling for mobile applications, offering an incremental improvement in efficiency for privacy-sensitive and latency-critical scenarios.
The paper tackles the challenge of efficiently running super-resolution models on resource-constrained mobile devices by proposing MobiSR, a framework that optimizes model compression and uses a scheduler to dispatch image patches based on difficulty, achieving average speedups of 2.13x and 4.79x over baseline implementations.
In recent years, convolutional networks have demonstrated unprecedented performance in the image restoration task of super-resolution (SR). SR entails the upscaling of a single low-resolution image in order to meet application-specific image quality demands and plays a key role in mobile devices. To comply with privacy regulations and reduce the overhead of cloud computing, executing SR models locally on-device constitutes a key alternative approach. Nevertheless, the excessive compute and memory requirements of SR workloads pose a challenge in mapping SR networks on resource-constrained mobile platforms. This work presents MobiSR, a novel framework for performing efficient super-resolution on-device. Given a target mobile platform, the proposed framework considers popular model compression techniques and traverses the design space to reach the highest performing trade-off between image quality and processing speed. At run time, a novel scheduler dispatches incoming image patches to the appropriate model-engine pair based on the patch's estimated upscaling difficulty in order to meet the required image quality with minimum processing latency. Quantitative evaluation shows that the proposed framework yields on-device SR designs that achieve an average speedup of 2.13x over highly-optimized parallel difficulty-unaware mappings and 4.79x over highly-optimized single compute engine implementations.