A Novel Multi-Image Stripe Noise Removal Method Based on Wavelet and Recurrent Networks

The stripe noise present in ultraviolet (UV) images often stems from the nonuniformity of the detector's readout circuit, posing challenges for subsequent image processing tasks. Existing destriping algorithms typically rely on single-image methods, failing to fully leverage the correlation across multiple consecutive images. This article proposes a novel multi-image wavelet deep recurrent network (MWDRN) to effectively remove the stripe noise from images. By inputting multiple stripe images with pixel shifts, MWDRN capitalizes on the complementary information present across these images to effectively remove the stripe while preserving genuine information. Initially, the high-frequency coefficients of Haar discrete wavelet transform (HDWT) are harnessed to target stripe. The encoder is tasked with extracting features from the multiple images, while a convolutional gated recurrent unit (ConvGRU) establishes the correlation among these features, ensuring the preservation of image details. In the proposed method, the Shiftnet is employed to align pixel-off images, and the decoder generates a destriping image in the final layer. The experimental results, both on UV images and public datasets, demonstrate the superiority of the proposed method over several state-of-the-art stripe noise removal methods.