DCT-based Algorithm for Rain Streaks Removal in Still Images

Rain Streaks Removal (RSR) for video and still images helps improve picture quality and facilitates subsequent image processing, and is therefore a popular and challenging research topic recently. Reviewing the literature, we found that machine learning (ML) based algorithms have been widely used for rain streaks removal due to the advances of artificial intelligence (AI) technology in recent years. Other than ML-based approaches, a discrete cosine transformation-based (DCT-based) algorithm for rain streaks in video / still image removal is proposed in this paper. There are two major parts in the proposed algorithm: the first part is to build a DCT coefficient-based rain streaks dictionary offline, the other is to perform rain streaks removal using the dictionary. With the nature of DCT, the texture characteristics of an image are decomposed into different coefficient patterns in frequency domain, and the DCT coefficients can be regarded as a feature space during the rain streaks removal process. To evaluate the performance of the proposed approach, the SPA-Data is used. Experimental results show that with the DCT-based feature space, the proposed algorithm not only makes rain streaks removal effective but also brings superior performance in various conditions, including different image resolutions and both real and synthesized rainy environments when compared with that of state-of-the-art approaches. Furthermore, due to the mature development of DCT technology, effective hardware and software instructions can be found in most of the computing platforms, which provides higher feasibility when the proposed algorithm is to be deployed on edge devices.