Qiyuan Guan

×10−61\times 10^{-6} using a cosine annealing scheduler. For the asymmetric MoE architecture, each expert is designed with a distinct self-attention capacity, enabling diverse feature modeling across varying attention complexities. We train the network for a total of 300,000 iterations. For multi-objective reweighted optimization, we set the step window size to 10 and the sensitivity parameter τ\tau to 5. 4.2 Comparison with the state-of-the-art We compare our proposed UniRain with 10 image deraining technologies, including PReNet [33], RCDNet [39], MPRNet [50], Restormer [51], IDT [44], DRSformer [5], RLP [53], MSDT [4], NeRD-Rain [7], and URIR [45]. Evaluation on the RainRAG dataset. Table 1 shows that the proposed UniRain achieves the best average performance across four rain degradation types. Compared with Restormer [51] and NeRD-Rain [7], it achieves improvements of 1.35 dB and 1.41 dB in PSNR on the DRD subset, respectively. As shown in Figure 3, we further present the visual comparisons under various rain degradations, where our method delivers superior deraining performance. In contrast, other methods still exhibit rain streak residuals and fail to recover background details affected by raindrops. Evaluation on real-world public benchmarks. For further general verification in practical applications, we conduct comparisons with other methods on the real-world rainy datasets. As shown in Table 2, our method achieves the best overall performance, outperforming URIR [45] by 1.73 dB in PSNR on average. Figure 4 shows that our method not only removes real and complex rain streaks but also restores clear details, even achieving visually better results than the GT by eliminating residual raindrops contained in it. Generalization to multiple application scenarios. We further evaluate the generalization ability of our method across multiple application scenarios, including autonomous driving, unmanned aerial vehicle (UAV), and maritime scenes. As presented in Figure 5, other methods struggle to preserve pixel-level structural fidelity, whereas our method produces faithful and visually pleasing results, demonstrating generalization across multiple scenarios. Model complexity. We evaluate the complexity of our method and state-of-the-art ones in terms of FLOPs and Params. As shown in Table 3, UniRain achieves competitive performance with lower FLOPs and fewer parameters. LQ Restormer [51] MSDT [4] NeRD-Rain [7] URIR [45] UniRain GT Figure 4: Visual comparison of image restoration results on the real-world benchmark (i.e., WeatherBench [14]). Compared to the state-of-the-art methods, our UniRain restores a high-quality image with clear details, even outperforming the GT by removing residual raindrops. Table 3: Comparisons of model complexity against state-of-the-art methods. The size of the test image is