Separated Frequency Loss with Distinguished Weight for Images Generation

Deep neural network models are more and more widely used in image reconstruction and generation tasks. By setting various loss functions, the model adaptively generates images that meet the corresponding constraints, which has a positive effect on improving image quality and enriching diversity. As an important image processing method, two-dimensional Fourier transform is widely used for image filtering and denoising, but its role has not been further explored. Based on the background of the fine-grained target image reconstruction task, this paper proposes a differentiated weight loss function based on frequency domain space in view of the problem that the global loss function of fixed constraints in the past brings great bias to model parameter update. The distinguishing regional weight coefficient is stetted by separating the high and low frequency regions, thus, to alleviate the high-frequency information introduced by the fine-grained target due to the diverse changes of individual morphology and environmental background. The model parameter update process gains much smoother with the guidance of separated frequency loss. Experimental results show that by setting the differentiated frequency domain region weight loss function, the model generates higher quality images in both image reconstruction and generation tasks.