Evaluation of Generative-AI Fashion Images using CNN Classification and Regression

The integration of artificial intelligence (AI) in the fashion industry has the potential to revolutionize design processes, enabling the generation of diverse designs based on user inputs. Despite the promise of AI, the unpredictable quality of outputs and the challenge of meeting human aesthetic and market standards remain significant hurdles. To address these issues, this research aims to (1) create a dataset of AI -generated fashion images labeled with customer rankings, and (2) establish a new evaluation method from the perspectives of the general public and the market. Data were collected from fashion e-commerce websites focusing on customer ratings and product images. Two CNN models were trained: one for regression, predicting continuous aesthetic scores, and another for classification, categorizing images into discrete rating intervals. Performance evaluation revealed that the regression model, constrained by clip techniques, maintained the original distribution of data while the classification model provided a benchmark for design indicators. The study concludes that addressing data imbalance and applying augmentation techniques yielded significant results. Specifically, the regression model achieved an RMSE of 0.866, while the classification model attained an accuracy of 93%.