Research On Clothes Design Algorithm Based On Style Transfer Neural Network

In the contemporary epoch of self-media,a substantial proportion of information consumed on the internet is derived from individual creators.Their content,typically imbued with creativity and distinctive personal flair,stands as a testament to the utility of self-media platforms.These platforms furnish creators with tools that substantially mitigate the complexity and professional demands inherent in content creation.In an endeavor to galvanize individual creators to partake in the fashion design industry and engender personalized clothing patterns,this thesis proffers a novel clothing design algorithm predicated on style transfer networks.Our proposed methodology rectifies the paucity of interactivity that plagues traditional style transfer algorithms by integrating an image aesthetics ranking network.This network provides consequential feedback scores on stylized images.In addition,an image matting network is instituted prior to the application of the style transfer algorithm,with the explicit aim of extracting the requisite foreground content from the user’s design material image.This enhancement amplifies the applicability of our algorithm within the context of creative clothing pattern creation scenarios.The principal aim of the methodology delineated in this thesis is to leverage deep learning techniques to simplify the segments of the design process that demand a high level of professionalism.For individual creators,image materials are often derived from simple life photography,where the background can sometimes be overly complex.When only a specific portion of the material is required for style transfer,it necessitates foreground extraction.Accordingly,this thesis employs a deep image matting algorithm to effectively distinguish between the foreground and background of content images.Utilizing convolutional neural networks and trimap to calculate the alpha matte of the image,a satisfactory matting effect is achieved.Moreover,for images with a clear distinction between the foreground and background that contain common content,a semantic segmentation algorithm is utilized to obtain image masks.After undergoing erosion and dilation,these masks can automatically generate trimap devoid of labels,thereby substantially augmenting the practical flexibility of our method.Empirical evidence underlines the superior separation effect between the foreground and background engendered by the deep learning-based image matting algorithm.In a bid to generate personalized creative images,this thesis employs a style transfer neural network with an adjustable style intensity parameter to stylize content images,thereby bestowing upon these images abstract artistic style features while retaining elements of the original content information.A distinct network is trained for each specific style to ensure the swiftest response time.Subsequent to the style transfer neural network,an image aesthetics ranking network is connected.This network receives the output image of the style transfer network as input and yields an aesthetic quality score for the image,which is subsequently relayed back to the creator.This ranking network can also utilize the scores and labeled information of user-feedback works as a training dataset.Upon the accumulation of sufficient labeled data,the image aesthetic quality ranking network can assimilate the aesthetic proclivities of the general populace.Experimental results underscore the ability of the evaluation network to score a diverse array of images outputted by the style transfer neural network.These scores can be utilized to appraise the degree of congruence between each style and the current content image,and to compute the average aesthetic score of a specific style across various content images,thereby enriching the user’s interactive experience.