.png){kind=icon}
.png){kind=icon}
Return
Deep learning-based image enhancement trained on multiwavelength-weighted simulation data
.png){kind=icon}
PRE
AI
W
Z
L
L
W
L
王
DOI：10.1117/1.OE.65.4.043101.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deep learning-based image reconstruction for visible-near infrared cameras relies heavily on the quality and physical realism of simulated training data. However, most existing simulation dataset generation methods based on actual optical systems neglect the influence of multiwavelength weighting on imaging characteristics. To address this limitation, we propose a simulation data generation method that incorporates multiwavelength weighting variables. Based on the actual optical design, differentiable ray tracing is employed to construct a multiwavelength-weighted training data, and a standard nonlinear activation free network model is trained for image reconstruction. Experiments on real-world images demonstrate that the proposed strategy achieves superior restoration performance, yielding an average 18.02% reduction in the BRISQUE score, corresponding to an similar to 8% additional improvement compared with the model trained without multiwavelength weighting. These results indicate that incorporating multiwavelength weighting into simulation data generation is effective and feasible for enhancing image reconstruction performance in real-world environments.
Keywords:
image enhancement
differentiable ray tracing
dataset enhancement
simulation data generation
Journal
O
IF:
1.2
Papers:
178
Citations:
1.1W
Organization
Cited Papers