安徽大学 电子信息工程学院，安徽 合肥 230601

足迹图像作为刑侦与生物识别中的关键个体特征，在采集过程中易受多种环境因素干扰，常伴随复杂噪声与图像质量下降。针对足迹图像中常见的复合噪声，本文提出了一种改进的双分支循环去噪网络，实现了高保真图像还原与纹理结构重建。网络整体包含两个生成器与两个判别器，生成器部分由两个协同优化的分支组成：去噪映射分支与颜色校正分支。其中，去噪映射分支设计了增强型多尺度结构块（EMSB）以强化结构建模与纹理恢复能力，融合多尺度卷积、深度可分离卷积与多注意力机制，有效增强纹理敏感区域的特征表达能力；颜色校正分支构建了自适应颜色一致性模块（CCM），通过多尺度残差卷积提取颜色特征，并在RGB空间中进行通道归一化与残差融合，以抑制生成图像中的色偏问题。此外，本文设计了多层级结构感知损失函数（MSSP-Loss），联合像素精度与结构相似性，引导网络在恢复细节的同时提升整体感知质量，并在自建足迹数据集FSD-Real上开展了实验评估。结果表明，本文方法在PSNR与SSIM指标上分别达到30.3dB和0.926，显著优于现有主流方法。同时，在主观视觉效果上亦展现出更强的去噪能力与细节保留能力，验证了其在实际足迹图像处理任务中的应用潜力。

As a critical biometric feature in criminal investigation and biometric recognition, footprint images are highly susceptible to various environmental interferences during acquisition, often accompanied by complex noise and degradation in image quality. To address the prevalent compound noise in footprint images, this paper proposes an improved dual-branch cyclic denoising network capable of high-fidelity image restoration and texture structure reconstruction. The overall framework consists of two generators and two discriminators. The generator is composed of two synergistically optimized branches: a denoising mapping branch and a color correction branch. Specifically, the denoising mapping branch incorporates an Enhanced Multi-Scale Structure Block (EMSB) to strengthen structural modeling and texture recovery. This module integrates multi-scale convolution, depthwise separable convolution, and multiple attention mechanisms to enhance feature representation in texture-sensitive regions. The color correction branch introduces an adaptive Color Consistency Module (CCM), which extracts color features through multi-scale residual convolution and performs channel-wise normalization and residual fusion in the RGB space to suppress color distortion in generated images. Furthermore, a Multi-Stage Structure-Perception Loss (MSSP-Loss) is designed, combining pixel-level accuracy with structural similarity to guide the network in recovering fine details while enhancing perceptual quality. Experiments conducted on the self-constructed footprint dataset FSD-Real demonstrate that the proposed method achieves a PSNR of 30.3 dB and an SSIM of 0.926, outperforming existing state-of-the-art methods. In addition, the method shows superior denoising performance and detail preservation in subjective visual evaluations, validating its potential for practical applications in real-world footprint image processing tasks.