Research On Image Style Transfer Method Based On Foreground And Background Separation

Image style transfer can be used in art creation,film and television special effects,etc.It is a research hotspot and difficulty in the current image processing field.With the rapid development of the mobile Internet,it is popular to create fantasy pictures through style transfer and share them on social networking sites.However,the current style transfer is mainly aimed at the entire image,lacking variability.In order to enrich the diversity of style transfer and provide new creative methods for people,this paper mainly studies the image style transfer technology with different foreground and background,and finally designs and implements an image style transfer system with interactive foreground and background separation.The main research results of this paper are as follows:(1)Aiming at the problems of automatic segmentation and poor segmentation accuracy of foreground person objects,we studied and implemented the foreground objects segmentation method based on region recommendation.Firstly,constructing the Mask R-CNN network structure.And then using the ResNet and the feature pyramid network to extract multi-scale features.Secondly,using the region proposal network to generate the anchor box dynamically and select the candidate region.Thirdly,the region of interest in the feature map is aligned with the region of interest in the input image.Finally,achieving foreground object segmentation by the detection,classification and segmentation task regression.(2)For the problems of lack of color and texture features in current style transfer,we proposed a style transfer method based on convolutional neural network.Training the simplified VGG-19 model,establishing a network structure for calculating and optimize the content and style loss function.Using the best feature map extracted by VGG-19 model to represent the content characteristics of the source image,and then using the Gram matrix to calculate each layer in the network and the relationship between the feature maps to be extracted is used to extract and represent the style features of the target image.The total loss function of the content and style features is continuously reduced by the Adam optimizer,and fit the image with the fusion of content and style to make the user satisfied.(3)Aiming at the problem that the foreground and background local style and the boundary is not smooth,we proposed a soft mask-based localized stylization method.Firstly,generating a soft mask between 0 and 1 by average filter and average pooling.And then adding the mask information to the Gram matrix to the contribution of each mask pixel to the Gram matrix.Finally,calculating the content and style loss respectively,and preventing gradient updating in non-stylized areas to discard edge noise and achieve boundary smooth.(4)Designing and implementing an image style transfer system with a foreground and background separation.In order to verify the feasibility and effectiveness of the proposed method,through detailed requirements research,functional modules such as self-selected images,foreground target segmentation,image style transfer,and human-computer interaction are realized.The system is stable in operation and works well.