LaMa network architecture search for image inpainting

Abstract

The neural architecture search problem is to obtain a neural network architecture with a version of the selected block that has the best performance according to a pre-selected evaluation strategy compared to other alternative versions. The aim of the paper is to improve the performance of image inpainting using neural architecture search by applying the wavelet transform to the LaMa network. Analyzing the results of experiments on researching the performance of image inpainting using the developed software it was noticed that the inpainting was better for images containing significant areas of uniform intensity, finegrained or structural texture. Fragments of images, including complex textures or detailed patterns were inpainted worse. The proposed technique for searching neural architecture for image inpainting based on LaMa differs in the ratio of image inpainting time and the quality of the reconstructed image. Inpainting of images with large masks based on the LaMa network is improved by applying the wavelet transform. In particular, the quality of filling the missing areas with image edges and small details is improved. In addition, it was researched the dependence of the quality of generating of details and edges of objects in the image on the properties of the image textures, which can be described by texture descriptors. Prospect for further research is prediction the effectiveness of the image inpainting with the LaMa networks depending on the estimated values of original image texture descriptors and missing areas size.