Image deraining via multi-level decomposition and empirical wavelet transform

Abstract

Image deraining, a crucial process in image restoration, finds wide-ranging applications in computer vision. Existing state-of-the-art deraining techniques, predominantly relying on image smoothing, dictionary learning, sparse coding, and deep neural networks, often fall short in delivering desirable outputs when faced with heavy rain. In this research article, we propose an advanced approach for image deraining, employing a multilayer decomposition strategy based on Empirical Wavelet Transform (EWT) and Dual Dictionary Learning (DDL). The proposed method introduces the Dark Channel Prior (DCP) in the preprocessing stage and utilizes Frequency Discrimination (FD), Empirical Wavelet Transform, and sparse-based methods with Dual Dictionary Learning to generate one low and three high frequency (HF) decomposed image components. The rain parts are subsequently removed from each HF image component through morphological decomposition in multiple layers. The non-rain outputs are combined with the lower frequency image obtained from the bilateral filter output to produce the rain-free image. The final output is further refined by adjusting the contrast, sharpness, and color balance of the de-rained image. To validate the efficacy of our proposed algorithm, we conducted a comprehensive evaluation using both subjective (visual quality) and objective (quantitative quality metrics) approaches. Comparative analysis with state-of-the-art methods confirms that our method outperforms existing techniques, demonstrating superior image-deraining capabilities. The proposed approach showcases promising results in addressing the challenges posed by heavy rain, establishing it as a robust and effective solution for image-deraining applications in various computer vision domains.