Improved high dynamic range imaging using multi-scale feature flows balanced between task-orientedness and accuracy

• DOI: 10.1016/j.cviu.2024.104126
• Publication Date: 2024-08-24T19:00:48Z
• Authors: Qian Ye; Masanori Suganuma; Takayuki Okatani
• Keywords: Feature (linguistics)

Deep learning has made it possible to accurately generate high dynamic range (HDR) images from multiple taken at different exposure settings, largely owing advancements in neural network design. However, generating without artifacts remains difficult, especially scenes with moving objects. In such cases, issues like color distortion, geometric misalignment, or ghosting can appear. Current state-of-the-art designs address this by estimating the optical flow between input align them better. The parameters for estimation are learned through primary goal, producing high-quality HDR images. we find that "task-oriented flow" approach its drawbacks, minimizing artifacts. To this, introduce a new design and training method improve accuracy of estimation. This aims strike balance task-oriented accurate flow. Additionally, utilizes multi-scale features extracted both image reconstruction. Our experiments demonstrate these two innovations result fewer enhanced quality.