<p>Single-image deraining is a critical preprocessing step for many outdoor computer vision applications, as rain streaks can obscure important landscape details and degrade visibility. Since rain streaks and image details are often mixed together in rainy images, and since they interfere with each other during the rain removal process, it is important to select useful information for accurate estimation of rain streaks. We propose a bright feature selection transformer (BFCformer) to effectively learn rain streak features. Our work is motivated by the observations that the features of rain streaks are generally brighter than those of surrounding areas without rain, that they are different from the rain-free features, and that the rain streak features in different regions are similar. This understanding inspired us to develop a bright feature selection self-attention (BFSSA) that estimates the rain streak features in two steps: First, the bright anchor points containing the rain streak information are filtered from the original features, and then, the rain streak features that are similar to the filtered anchor points are selected from the original features. Simultaneously, to determine which low- and high-frequency information should be preserved in both the local and global domains, we developed a dual-branch discriminative frequency domain-based feed-forward network (DDFFN) that aggregates useful features in both the local and global branches. Extensive experiments show that the proposed BFCformer outperforms other state-of-the-art single-image deraining methods on synthetic datasets and real-world rainy scenes. The source code and pretrained models are available at <a href="https://github.com/Lvya123/BFCformer">https://github.com/Lvya123/BFCformer</a>.</p>