| Autor: |
Ma, Lifeng, Hu, Yanxiang, Zhang, Bo, Li, Jiaqi, Chen, Zhijie, Sun, Wenhao |
| Předmět: |
|
| Zdroj: |
Applied Intelligence; Jan2023, Vol. 53 Issue 2, p1452-1468, 17p |
| Abstrakt: |
With a span of last several years, Deep Learning (DL) has achieved great success in image fusion. For Multi-Focus Image Fusion (MFIF) task, focus classification learning based methods are the most popular ones. This type of methods seeks to generate an all-in-focus synthetic image by combining the partial focused source images according to their focus properties. The basic premise they rely on is that the fused sources are focus complementary. However, this two-classification model is not always valid in practice and consequently, leads to the quality degradation in the focus/defocus junction regions. In addition, the widely used single-scale stitching rule makes them lack the robustness to the source misregistration, which is hard to avoid in the practical use. To address these drawbacks effectively, we propose a new multi-classification focus model and multi-scale decomposition based MFIF method, termed as MCMSCNN, in this paper. Concretely, we design and train a CNN classifier to obtain an initial relative focus probability map of the sources at first, and then we fuse the sources within a multi-classification and multi-scale decomposition based fusion framework. In this framework, we propose a multi-classification focus model to characterize the pixel focus property and fuse various categories of pixels with more specific rule in a multi-scale manner. All these means make our method presents more outstanding focus fusion performance and anti-misregistration capability. In the experiments, we contrast our method with some recently proposed MFIF methods by subjective and objective comparisons. Extensive experimental results validate the competitive performance of the proposed method. [ABSTRACT FROM AUTHOR] |
| Databáze: |
Complementary Index |
| Externí odkaz: |
|
Full text from SpringerLink