Research On The Techniques Of Image Appearance Transfer Based On Multi-feature

Image appearance transfer is the process of transferring the color,texture,or illumination features of the reference image to the target image,which can change the appearance of the target image and meet user's requirements.Appearance transfer provides users a simple way to make image editing.However,traditional methods cannot meet all the requirements.On the one hand,traditional methods are usually based on a single feature,such as color feature or texture feature,and cannot get sufficient transfer results.On the other hand,when the illumination feature becomes complex,traditional appearance transfer method cannot get satisfactory results.Therefore,we proposed new appearance transfer method based on multi-features for each of the two problems separately.Firstly,in order to change the color and texture features of the image at the same time,we present an appearance transfer algorithm that combines color feature and texture feature.Taking advantage of the feature extraction and matching,this method can automatically detect the regions that need to do color transfer and texture transfer.For the sake of being more flexible for appearance transfer,this thesis also proposes a stroke-based interactive system.User can specify the regions by strokes.Then,color transfer and texture transfer are performed between the corresponding regions.Finally,change the color and texture information of the image at the same time.Secondly,in order to overcome the problems caused by the differences of illumination features between images in the existing methods,this thesis designs and implements an image appearance transfer method under complicated illumination.Since the traditional methods are usually a linear transfer,it is difficult to deal with the non-linear transformation.A convolutional neural network is introduced to learn the illumination mapping between target image and reference image under complex illumination.In order to refine the appearance transfer result,we adopt the joint bilateral filter to remove the noise generated after the calculation process of the full convolution network by using the input image as the guide image.The results show that the method can solve the above problem.