Yang Yutian, Lin Jianyu, Dai Xinyue, Zhang Zhipei, Zhang Shuijin, Chen Yingyu, Kong Guangxin, Li Xin
College of Computer Science and Technolog, Civil Aviation University of China, Tianjin, China.
University of Electronic Science and Technology of China, Zhongshan Institute, Zhongshan, China.
PLoS One. 2025 May 16;20(5):e0322011. doi: 10.1371/journal.pone.0322011. eCollection 2025.
In this paper, a novel hybrid network called ChessFormer is proposed for the single image de-rain task. The network seamlessly integrates the advantages of Transformer and fitted neural network (CNN) in a checkerboard architecture, fully utilizing the global modeling capability of Transformer and the local feature extraction efficiency of CNN.ChessFormer adopts a multilevel feature extraction and progressive feature fusion strategy to efficiently achieve the rain line while preserving the We design a multidimensional transposed attention (MSTA), which enhances the network fusion for different rain patterns and mechanism image textures by combining self-attention with gated phase operation. In addition, the efficient architecture ensures full integration of features across dimensions and codecs. Experimental results show that ChessFormer outperforms existing methods in terms of quantitative metrics and visual quality on multiple benchmark datasets, achieving state-of-the-art performance with fewer parameters.
本文提出了一种名为ChessFormer的新型混合网络用于单图像去雨任务。该网络在棋盘架构中无缝集成了Transformer和卷积神经网络（CNN）的优势，充分利用了Transformer的全局建模能力和CNN的局部特征提取效率。ChessFormer采用多级特征提取和渐进式特征融合策略，在保留……的同时有效地实现雨线。我们设计了一种多维转置注意力（MSTA），通过将自注意力与门控相位操作相结合，增强了网络对不同雨型和机制图像纹理的融合。此外，高效的架构确保了跨维度和编解码器的特征完全集成。实验结果表明，ChessFormer在多个基准数据集上的定量指标和视觉质量方面优于现有方法，以更少的参数实现了当前最优性能。
