A Vignetting-Correction-Based Underwater Image Enhancement Method for AUV With Artificial Light

Abstract

Images captured by autonomous underwater vehicles (AUVs) are inherently affected by artificial light, which tends to generate distinctive footprint and biased veiling light on the foreground. Existing underwater image enhancement (UIE) methods do not take this serious problem into account. In practice, the enhanced results will lead to a severe performance drop, due to the challenging joint task of enhancing underwater images while correcting the vignetting phenomenon. To solve this issue, we propose a two-stage vignetting-correction driven UIE network (called VCU-Net), which consists of two subnetworks (vignetting-correction-net and restoration-net), to deal with the two joint tasks in a split way. Concretely, we first introduce a novel underwater imaging model that is more capable of describing the imaging process for underwater robot applications. Accordingly, sufficient underwater data with vignetting is conducted to train our VCU-Net. In addition, based on the intensity distribution statistics of the lighting footprint formed by artificial light, a radial gradient constrained loss is designed in the vignetting-correction-net, which facilitates the precise estimation of vignetting. To validate the performance, extensive experiments on both synthetic and real-world images captured with AUV show the effectiveness of the proposed novel method, which illustrates a great superiority against the state-of-the-art methods in real underwater world with complex illumination.