Method And Research On Style Transfer To Keep Target Main Body Invariant In Deep Learning

Researchers and practical application personnel see new opportunities and big challenges in the field of artificial intelligence,and arouse the attention.Image problem is a hot-point topic in the field of artificial intelligence.Where,image style transfer and image segmentation are one core of image problems.They are also key topics studied in computer vision.Most of the images are composed of the target main body and the background.The target main body is important information which presented to people;The background is specific situation and affects the important factors of the target main body.Traditional image style transfer method extract styles from one image,and then apply all extracted styles to another image for newly perceived image.How to extract the target main body and transfer the background style to produce different perception effects is the problem to be solved in this paper.Based on deep learning,this paper uses image segmentation techniques such as extraction,detection,recognition,etc.,to obtain the target main body.On the basis of retaining the image target main body,the image style transfer technique applies to transfer the background,thereby generating new perceive images and new perception effects.This paper proposes a new method that combines image segmentation and image style transfer to enhance the image perception effect and highlight the target main body.Using semantic cascade Mask R-CNN method explore more image information to achieve better image segmentation effect.Semantic cascading Mask R-CNN processes the image to obtain the target main body and backbone network consists of the 101-layer convolutional neural network,residual network(Res Net)and feature pyramid(FPN);it integrates cascade into target main body by interweaving detection and segmentation features together for a joint multi-stage processing;it adopts a fully convolutional branch obtain spatial context to present the semantic information of the image,which can help distinguishing hard target main body from the background.Using image style transfer method(i.e.,Neural Style Transfer)enhance the image perception effect.This method uses convolutional neural networks(CNN)to open up a new chapter in image style transfer and uses vgg19 as the backbone network;it can separate and reorganize image content and natural image styles.In addition,attractive new styles are generated from existing styles by decomposing styles into perceived factors to combine style information from multiple sources.Experimental results show that the method proposed in this paper produces different perceived effects to other recent methods.The main contributions of this paper are listed as follows.1.A method of combining image segmentation and image style transfer is proposed,which can enhance the image perception effect,highlight the target main body,and solve the problem that the traditional pattern transfer method does not consider the original state of the target main body.2.Using the semantic cascade Mask R-CNN method explore more image information,distinguish between hard target main body and background,and retain more image target main body features,to solve the problem that Mask R-CNN does not consider semantic information and target main body features in detection function.3.The image styles are separated and reorganized through the style transfer method,only choose to transfer the background styles,to solve the limitation of the traditional style transfer,all the extracted styles can only be transferred to another image as a whole.