Estimation of Garment Impression using Regression Models with Design Parameters and Image Features

Abstract

To estimate the garment impression, we verified three regression models using design parameters, geometrical and texture features of images, and convolutional features of images. Adopting three-dimensional apparel simulation, we generated 375 images of a men's outdoor jacket by changing design parameters, namely the length, waist, hem circumference, and sleeve circumference. Nine participants evaluated the cool-uncool (kakkoī-kakkowarui in Japanese) impression of the garment images using a semantic differential method. We used the measured widths, angles, and G-type Fourier descriptors of the jacket images as geometrical features. For texture features, the fractal dimension and Haralick texture features (Asm, Entropy, Contrast, Correlation) of the images were used. Convolutional features were obtained from the images using a convolutional neural network. By using the design parameters and the obtained features, we estimated the image impression using three regression models: multiple linear regression, neural network, and light gradient boosting machine (LightGBM) models. As a result, the LightGBM with the design parameters had the highest mean correlation coefficient for all participants. The inclusion of design parameters was thus found to be effective in estimating the impression based on garment design parameters with LightGBM.