Double iterative optimal dictionary learning-based SAR image filtering method

Abstract

The KSVD algorithm has been a focus of image filtering research ever since it was first developed. However, this algorithm has a slow convergence rate for sparse coding, and dictionary updating via singular value decomposition is too complex. Consequently, the filtering efficiency is very low for large synthetic aperture radar images, and a significant amount of edge texture information is lost. To address these issues, this paper proposes a double iterative optimal dictionary (DIOD) learning algorithm. First, during the sparse coding process, a single iteration is performed to select the optimal and second most optimal dictionary atoms for representing the residual error, and the selected atoms are then updated according to the accumulation of dictionary coefficients. Next, the normalized and weighted reconstruction error is used to update the dictionary. The experimental results indicate that the DIOD algorithm provides better edge preservation and has better operational efficiency than the KSVD algorithm while maintaining a high peak-signal-to-noise ratio. This approach effectively avoids the repetitive occurrence of dictionary atoms in the iteration process, accelerates the convergence rate, and increases the sparsity of the sparse coding coefficients while simplifying the dictionary updating method, enabling rapid filtering of large synthetic aperture radar images, and achieving an improved filtering effect.