Abstract: To address the issue of low binding force of the Fréchet Inception Distance (FID) metric in cross-domain image generation evaluation,a domain difference calculation method based on the Multi-Level Feature Alignment Distance (MLFAD) is proposed.By introducing a multi-level feature extraction and feature alignment mechanism,a comprehensive measurement of the distribution differences of generated images and real images at different hierarchical feature scales is achieved,which more accurately reflects the detailed and global information of the images and significantly enhances the binding force on the generated images.The experimental results show that the performance of MLFAD is significantly better than that of the FID.In cross-domain image generation tasks such as image translation,style transfer,image restoration,and image enhancement,the domain difference has been reduced by 31.7%,26.8%,41.1%,and 34.6%,respectively.Therefore,MLFAD which has excellent accuracy and stability in the evaluation of cross-domain image generation tasks,significantly improves the clarity,detail restoration,and color accuracy of the generated images.

Key words: Image processing, Image quality assessment, Feature extraction, Feature fusion