画像変換AIにおける画像の情報量に依る生成画像精度の非対称性について

Abstract

In this study, we demonstrated the asymmetry between the concretization task and the abstraction task by pix2pix, which is the basis of current image transformation AI. For the AtoB model, which converts satellite images to map images, and the BtoA model, which converts map images to satellite images, the entropy of each image was calculated after obtaining the GLCM of each image, and the correlation between the entropy ratio of input image to output image and the accuracy of the generated image by pix2pix was obtained. The results showed that there is a negative correlation in the AtoB model, which is responsible for the abstraction task, where the accuracy decreases as the image entropy ratio increases, and only a weak negative correlation in the BtoA model, which is responsible for the concretization task. Then it has been demonstrated that pix2pix have an asymmetric property between abstract tasks than concrete tasks. We investigated this property by focusing on the L1 error of U-Net, which is frequently used in image generation models, and found that the convergence speed in the AtoB model is faster than in the BtoA model. From this, we infer that the solution space of the weights is sparser in the abstraction task, and that learning proceeds toward larger amounts of entropy due in part to the skip-connection effect of the U-Net structure. In the future, as the characteristics of U-Net become clearer, it can be used to improve the accuracy of image abstraction and lower dimensionality.