Aiming at the complex and diverse interference of image moir é patterns, the difficulty in fully preserving texture details and edge information, and the poor performance and high computational complexity of existing moir é pattern removal models, a new moir é pattern removal network model based on UNet network is proposed, which integrates weighted convolution and multi-scale attention mechanism. In the feature extraction stage, a weighted convolutional layer based on density parameters is introduced to enhance the central region response of the convolutional kernel through adaptive weight adjustment, while suppressing the influence of edge noise. Secondly, in order to more effectively capture multi-scale feature information, spatial attention blocks and multi-scale channel attention modules are used for deep level feature extraction to obtain more detailed and rich feature representations. In the decoding stage, a combination of skip connections and residual connections is used to enhance the robustness of the model through multi-scale output, and finally, Moir é fringes are removed through weighted convolutional layers. The experimental results show that the evaluation scores of the model on three public datasets, UHDM, TIP, and FHDMI, are better than those of mainstream models for removing image moir é patterns, proving that the model effectively improves the effectiveness of image moir é pattern removal. 针对图像摩尔纹干扰复杂多样、纹理细节与边缘信息难以被充分保留，以及现有去摩尔纹模型效果欠佳、计算复杂度较高等问题，本文提出了一种基于UNet的高效去摩尔纹算法MSFDNet（Multi-Scale Feature and Density-aware Network）。首先，在特征提取阶段引入基于密度参数的加权卷积层（WCL），通过自适应权重调整增强卷积核中心区域响应；其次，设计多尺度特征增强模块（MCFE），利用不同膨胀率卷积分支结合通道注意力机制，有效捕获跨尺度纹理特征；同时提出空间特征调制模块（SFM），通过空间注意力机制突出摩尔纹干扰集中区域，提升网络对空间分布差异的感知能力；在解码阶段结合跳跃连接与残差连接策略，实现多尺度信息的有效融合与鲁棒恢复。MSFDNet算法在TIP、FHDMi与UHDM三个公开数据集上，在PSNR、SSIM和LPIPS三项评估指标上均优于目前主流的去摩尔纹算法。在TIP数据集上，PSNR达到31.04 dB、SSIM达到0.943，较其他算法有明显提升；在FHDMi数据集上，PSNR达到24.695 dB，在UHDM数据集上SSIM提升至0.8154，均优于目前主流算法；所提算法能够在有效去除图像摩尔纹的同时保持图像纹理与边缘细节，兼具更强的泛化能力与鲁棒性。
Wei et al. (Tue,) studied this question.