Local image style transfer algorithm for personalized clothing customization design

Abstract

With the increasing demand for personalized clothing from consumers, the style transfer technology of clothing images has become a key link in clothing customization design. However, when transferring clothing image styles in complex backgrounds, problems such as poor local image style transfer and boundary artifacts often arise. To address these issues, an attention mechanism-based approach to local style transfer in cyclic generative adversarial networks has been proposed. By introducing attention mechanisms, more precise probability allocation has been achieved. In addition, this study designs a local artifact correction model based on an improved residual network. The experimental results showed that the proposed method had an average ratio of 0.83 for the best performance image in user perception evaluation, which was at least 23.9 % higher than other methods. In addition, the average distance similarity of this research method reached 0.244, which was at least 4.4 % higher than other methods. In terms of mean square error, the research method had a mean square error as low as 3426, which was at least 8.5 % lower than other algorithms. In addition, regarding the artifact correction part, the average opinion score of the proposed method was 2.9, which was at least 7.4 % higher than other algorithms. The mean square error of this algorithm was only 16.13, at least 34.2 % lower than other algorithms. This study verifies the effectiveness of the proposed method in local style transfer and artifact correction of clothing images, provides strong technical support for the field of clothing customization, and helps to promote technical progress in this field.