Abstract:
Recovering a high dynamic range (HDR) image from a single low dynamic range (LDR) image, namely inverse tone mapping (ITM), is challenging due to the lack of information in over- and under-exposed regions. Current methods focus exclusively on training high performing but computationally inefficient models, which in turn hinders deployment of ITM models in resource constrained environments and applications. In the present work, we propose a computationally efficient neural network with reduced latency and cost compared to state-of-the-art ITM models in order to employ ITM in environments with limited computing power.
To this end, we propose combining efficient operations with a novel mixed quantization scheme to produce a well performing but a computationally efficient mixed quantization network (MQN) which can perform single image ITM on a mobile platform. In the experimental analyses, ITM models trained using our proposed MQN perform on par with the state-of-the-art methods on benchmark datasets and obtain up to 10 times improvement on latency.
