GENERATION AND IDENTIFICATION OF COLOR TRANSFORMATION MODEL PARAMETERS FOR SPECIFIC OBJECTS USING SMALL-SCALE DATASETS

Abstract

 This study investigates learning methods for color transformation models of specific objects under challenging conditions with limited training data, focusing on applications in landscape architecture. Using a small dataset of pedestrian bridge images captured on-site and employing CycleGAN as the color transformation model, we conducted parallel survey research to examine the learning effectiveness of loss function weighting in specific object regions and methods for identifying effective model parameters during the training phase. The results demonstrate that loss function weighting successfully generates model parameters with superior transformation performance. Furthermore, we established that utilizing the Structural Similarity Index Measure (SSIM) as an image similarity metric, combined with the CUmulative SUM (CUSUM) method for change point detection, provides an effective approach for identifying optimal model parameters.