PoissonNet: Single Image Detail Enhancement Based on Poisson Regression
Jiang He, Ziqiang Wang, Ping Zheng, Haoxiang Zhang, Yuze Wang, Boming Song, Deqiang Cheng · 2023
Conventional image detail enhancement approaches commonly rely on filter-based or learning-based methodologies. In contrast, this study introduces a novel frequency domain model named PoissonNet. PoissonNet comprises two fundamental components: a sampling network and a weighting network. The sampling network employs linear sampling techniques to generate sampling features of multiple orders, which have been empirically validated as effective constituents for capturing detailed characteristics. The weighting network plays a crucial role in generating weight matrices to facilitate the effective fusion of sampled features from different orders. In this research, Poisson sampling coefficients are employed to perform regression fitting on the sampled features of each order. The optimal solution, namely the detail layer of the image, is obtained by constraining the objective function through a logarithmic maximum likelihood equation. Extensive experimental evaluations substantiate the remarkable generalization capabilities of the proposed image detail enhancement model. Moreover, the model provides a visually pleasing experience when evaluated on internationally recognized texture datasets.