基于BIFPN-GAN特征融合的图像修复算法研究

Research on Image Inpainting Algorithm Based on BIFPN-GAN Feature Fusion

为了使图像修复方法在结构重建的过程中实现结构与纹理信息之间的交互,提高修复的图像在语义上的真实性。在原有的双流生成网络基础上改进了一种基于BIFPN多尺度特征融合算法的双流结构图像修复网络。该网络采用耦合方式实现结构约束下的纹理合成与纹理引导下的结构重建,实现纹理与结构信息的有效利用,有利于生成语义更真实的图像。构建BIFPN多尺度特征融合网络,以实现重建、感知与风格损失的补偿,使融合后的图像实现全局的一致性。在训练阶段,采用了基于语义的联合损失函数,以增强图像在结构生成上的合理性。通过在CelebA和Places2数据集上与其他修复网络进行对比实验,证明了改进的图像修复方法的客观评价指标更优,更加有效地修复破损图像的结构和纹理信息,图像修复性能更优。

In order to enable the image inpainting method to achieve interaction between structure and texture information in the process of structure reconstruction and to improve the semantic realism of the repaired images,a two-stream structural image inpainting network based on BIFPN multi-scale feature fusion algorithm is improved on the basis of the original two-stream generative network.The improved network adopts a coupling approach to achieve texture synthesis under structural constraints and structure reconstruction under texture guidance,thus realizing the effective use of texture and structure information,which benefits the generation of semantically more realistic images.A BIFPN multi-scale feature fusion network is constructed to achieve the compensation for reconstruction,perception and style loss,so that the fused images can achieve global consistency.In the training phase,a semantic-based joint loss function is used to enhance the rationality of image on structural generation.Through comparative experiment with other inpainting networks on CelebA and Places2 datasets,it is demonstrated that the improved image inpainting method has better objective evaluation metrics,more effective restoration of structural and texture information of corrupted images,and better image inpainting performance.