| With the full opening of the new retail of apparel,the integrated marketing of online and offline multi-channels has become the mainstream trend,and the bulk and personalized generation of apparel advertisements is of great significance.In this paper,the intelligent generation technology based on generative adversarial network(GAN)is applied to the design of apparel print advertisements to realize the automatic synthesis of apparel product images and the layout generation of advertisements,thus reducing the manual production cost and improving the design efficiency of advertisements.Firstly,a conditional generative adversarial network(CGAN)was selected as the base model for the experiments to synthesize product images in print ads,and the experimental dataset was Fashion-MNIST.The input of the model added the conditional variable y to the original GAN,and the output was the clothing images of specified categories,mainly contained 10 categories such as T-shirts,jackets,skirts and pants.Compared the images synthesized by iterating the model to different epochs,the optimal number of iterations of the model was 60 epochs,which could synthesize a variety of garment images with complete and clear silhouettes,and the loss values of the generator and discriminator tended to be 0.680,and the total loss value of the model was 0.848.Secondly,in view of the problems of slow convergence of the loss function of CGAN and the need to improve the clarity of the synthesized images,the experiments introduced the Transformer module into the generator of CGAN to further construct the CT-GAN model,so as to enhance the ability of the model to extract image features.The results showed that the quality of the image synthesized by 10 epochs of the CT-GAN model was comparable to that of 20 epochs of the CGAN model,and when the epochs were increased to 60,the style diversity,silhouette integrity and image clarity of the CT-GAN synthesized garment image were significantly enhanced.At this time,the loss values of the generator and discriminator converge to 0.692,and the total loss value of the model was 0.730,which was 13.92% lower than that of the CGAN model,indicating that the learning ability of CT-GAN was significantly improved and the overall convergence performance of the model was better.The FL-GAN model was constructed to realize the layout generation of garment print advertisements,and the model consisted of a multimodal embedding network and a layout generation network.The data set of this paper was further obtained by collecting the layout data of clothing advertisements online and expanding them.The multimodal embedding network learned the image,text and attribute features in the layout and fused them into multimodal features y.The encoder in the layout generation network mapped the layout samples x to the feature space,and the generator mapped the random vector z to the layout space,which were trained against the discriminator.The model finally output an advertisement layout conposed of background,image,text,and title elements.In the subjective evaluation of image quality,the highest mean scores of multiple sets of layout images generated by FL-GAN were 7.88,8.38 and 7.65 in rationality,aesthetics and typographical neatness,respectively.In addition,the objective evaluation results indicated that the overall quality of the layouts generated by the FL-GAN model was excellent.The quality of the images generated by the above models was subjectively evaluated by the expert scoring method,and the distortion of the images was calculated by using PSNR and SSIM objective evaluation indexes,and the results showed that the images generated by CGAN,CT-GAN and FL-GAN were reasonable and beautiful,and the quality of the images synthesized by the improved model CT-GAN on the basis of CGAN was significantly improved,as shown by The corresponding scores of style diversity,contour integrity and sharpness indexes were improved by 1.98%,3.43% and 8.54%.The GAN-based image synthesis and layout generation model built in this research can quickly realize the synthesis of various styles and types of clothing advertising product images and the generation of advertisement layouts including images,headlines,texts,backgrounds and other elements for clothing merchants to select and use,providing a digital solution for intelligent marketing and realizing cost reduction and efficiency increase in the process of advertisement production under the premise of batch and personalized generation. |
