Computational model for predicting user aesthetic preference for GUI using DCNNs

Abstract

Visual aesthetics is vital in determining the usability of the graphical user interface (GUI). It can strengthen the competitiveness of interactive online applications. Human aesthetic preferences for GUI are implicit and linked to various aspects of perception. In this study, an aesthetic GUI image database was constructed with 38,423 design works collected from Huaban.com, a popular social network website for art and design sharing, collection, and exhibition in China. The numbers of user collection and likes of each design work were used as the annotation to represent user preference levels. Deep convolutional neural networks were applied to evaluate the aesthetic preferences of GUIs, based on a large dataset of user interface design images with the ground-truth annotations. The experimental result indicated the feasibility of the proposed method, with a mean squared error (MSE) of 0.0222 for user collection prediction and an MSE of 0.0644 for user likes prediction in the best model performance of Squeeze-and-Excitation-VGG19 networks (SE-VGG19). This study aims to build a large aesthetic image database, and to explore a practical and objective evaluation model of GUI aesthetics.

A Deep Learning Model for the Assessment of the Visual Aesthetics of Mobile User Interfaces

“…With the advances in machine learning, deep learning approaches have been introduced to predict GUI visual aesthetics [Dou et al, 2019;Khani et al, 2016;Xing et al, 2021]. Some approaches employ pre-trained networks to reduce their training effort.…”

“…Instead of using pre-trained networks, Xing et al [2021] trained five distinct models with 38,423 GUI images collected from a popular website for UI designers in China. The dataset has been labeled using the number of "likes" the GUIs have received and the number of user "collections" to which they belong as their visual aesthetics representations.…”

“…For example, "2" is closer to "3" than it is to "5," but the cross-entropy loss would be the same. Also, regression has achieved better results than classification in similar works [Dou et al, 2019;Xing et al, 2021]. That way, the output is a numerical value within [0..1], interpreted as the visual aesthetics degree, where 0 = "very ugly" and 1 = "very beautiful.…”

“…More recently, deep learning models have been applied to quantify the visual aesthetics of the design of GUIs of webpages [Dou et al, 2019;Khani et al, 2016;Xing et al, 2021] following the success of the evaluation of the aesthetics of photographs [Deng et al, 2017;Lu et al, 2014;Malu et al, 2017]. Deep learning approaches can automatically extract high-level features directly from raw input data to predict the visual aesthetics of images [LeCun et al, 2015;Polyzotis et al, 2017].…”

A Study of Sustainable Product Design Evaluation Based on the Analytic Hierarchy Process and Deep Residual Networks

“…This approach combined the benefits of both methods, providing a comprehensive evaluation that was both objective and captured the human evaluators' expertise. Xing et al [28] constructed a dataset using GUI images with the ratio of likes/views as the ground-truth annotation, and proposed a GUI (graphic user interface) aesthetic evaluation model based on squeeze-and-excitation VGG19 network architecture. Zhang et al [29] constructed a self-adaptive deep aesthetic model of Chinese ink paintings using a subject query mechanism.…”