A Spatial Fusion Scheme of Multi-focus Image Combining SVM-Based Classification and PCA-Based Weight

The image fusion technique aims to generate a comprehensive image that can combine complementary information from different source images. Traditional signal processing-based image fusion methods are well-studied in recent years, and researchers try to explore new ideas for this task. Especially, the machine learning method likes support vector machine (SVM) has an apparent advantage in many situations. In this work, a sliding window technique is first used to extract the key features of source images based on several effective evaluation metrics which can reflect the detailed information. Second, the extracted image features are employed to distinguish the focused areas and unfocused areas of source images by a trained SVM model, so the decision results for each source image will be obtained. Third, consistency verification (CV) is utilized to assimilate a single singular point of a specific region in the decision results in order to correct possible errors. At last, a new weighted fusion method for the pixel is designed based on principal component analysis (PCA) to deal with the disputed areas which come from the same position of decision maps of different source images. Experimental results reveal the proposed method is effective and can achieve better performance than comparative methods.