A Two-stage Deep Network for High Dynamic Range Image Reconstruction

Sharif, S. M. A.; Biswas, Mithun; Kim, Sungjun

doi:10.48550/arxiv.2104.09386

Cited by 1 publication

(3 citation statements)

References 41 publications

(63 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Comparison of the latency of the models on the GPU platform is not fair, since our MQN model uses 8-bit integer quantization and employs depthwise convolutions which are not suited for GPU platforms [49]. To obtain a fair comparison on the mobile deployment platform, we adapt the four methods HDRUNet [71], ExpandNet [23], DeepHDR [72] and TwoStage [35] that perform the fastest inference on the GPU. The results show that the difference between the latency of our MQN models and these four models increases further on the mobile platform.…”

Section: Comparison With State-of-the-art Methodsmentioning

confidence: 99%

“…These facts, along with learning representations of HDR content over the input, enables us to obtain a faster model with similar accuracy. However, other models use methods that hinder efficient deployment, such as by utilising images with same resolution [23,35], inefficient network composition [24], convolution operations and special blocks [71,72].…”

Section: Comparison With State-of-the-art Methodsmentioning

confidence: 99%

“…The methods in the first group [19,23,24,26,29,30,31,32,33,34,35,36,37,38] suffer from the problem of not accurately processing over-and under-exposed regions, but provide more compact systems. Methods in the second group use a group of bracketed overand under-exposed images as input to directly learn to generate HDR output [39,39,40,41,42,43,44,45,46], as similarly utilized in photographic HDR generation.…”

Section: Related Workmentioning

confidence: 99%

See 2 more Smart Citations

A Mixed Quantization Network for Computationally Efficient Mobile Inverse Tone Mapping

Borrego-Carazo¹,

Özay²,

Laboyrie³

et al. 2022

Preprint

View full text Add to dashboard Cite

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 ITM models, which in turn hinder deployment of the ITM models in resource-constrained environments with limited computing power such as edge and mobile device applications.To this end, we propose combining efficient operations of deep neural networks with a novel mixed quantization scheme to construct a well performing but computationally efficient mixed quantization network (MQN) which can perform single image ITM on mobile platforms. In the ablation studies, we explore the effect of using different attention mechanisms, quantization schemes and loss functions on the performance of MQN in ITM tasks. In the comparative analyses, ITM models trained using MQN perform on par with the state-of-the-art methods on benchmark datasets. MQN models provide up to 10 times improvement on latency and 25 times improvement on memory consumption.

show abstract