Abstract:
In complex underwater environments, the image quality will be damaged due to light absorption, water quality and other factors, which brings certain challenges to the und...Show MoreMetadata
Abstract:
In complex underwater environments, the image quality will be damaged due to light absorption, water quality and other factors, which brings certain challenges to the underwater object detection task. In order to improve the quality of under-water images, as well as the effectiveness of underwater detection tasks, we proposed a multi-reference mapping based image enhancement network (MRUIE). Previous image enhancement methods set only one reference image for the original image to be learned, ignoring the diversity of real underwater images. We optimized for the salient problems of underwater images by allowing the network to learn multiple reference images to capture the uncertainty of real underwater images. First, for the prominent problems of color distortion, contrast imbalance, and low brightness and darkness in underwater images, we designed three optimization paths to generate three reference images respectively. Then, the three reference images are fed into the feature extraction network to construct the statistical distribution of features and generate a series of potential enhancement distributions. Finally, based on the enhancement distributions, Monte Carlo likelihood estimation is used to determine the final enhancement results. Experiments conducted on two datasets demonstrate that our proposed image enhancement algorithm can effectively enhance the restoration of the original underwater images and provide significant performance improvement in underwater vision tasks.
Date of Conference: 30 June 2024 - 05 July 2024
Date Added to IEEE Xplore: 09 September 2024
ISBN Information: