Local Color Transfer Based On Region Matching

As the increasing use of digital cameras, People begin to record the number of meaningful, memorable scenes in the tourism, parties and other important occasions. However, as the scene of the venue or the weather or light is unsuitable, the camera performance limitations and the limitations of the user's own photography and other reasons, many of the photo shoot the color will appear a variety of problems, Such as low (or over) exposure, some of the color deviation and other issues, since digital cameras and storage devices are fast and convenient, people tend to take large number of photos with the same or similar scenes at the same time, these photos can always find a satisfactory photo effects, therefore, based on the above situation, it is meaningful to provide a simple and effective color transfer algorithm.Color transfer is an image processing technique to enhance the color characteristics of a defect image (source image) by referring to a good image (sample image).Based on the existed algorithms, the paper proposes two improved local color transfer methods based on given sample image, they are local color transfer based on image character and local color transfer based on image understanding. For a source image that needs to be changed partially, users need to mark the region that they want to improve and the region to be preserved on the source image with strokes, and they also need to mark the correspondent reference region on the sample image with the strokes of the same color. Based on the strokes of the two images, region needed to be modified on the source image and the region used as reference on the sample image are obtained by employing the region extraction method. As for the images needed to be changed globally, the mentioned operations above are omitted. Unsupervised image segmentation method is utilized to segment the two images into several meaningful regions respectively according to different resolution images. For the first method image pyramid of both images are constructed, and based on the affiliation relationship between different levels of regions and color and texture features, characteristics of the similarity calculation region, the correspondent relationship between the two images are figured out automatically; Method two finds the corresponding relationship by employing extra image database. Finally, according to the correspondence between regions, the Gaussian-based statistical local color transfer method is applied to modify the color of the source image.The main research work and major contributions of our work are:1) Studying systematically recent color transfer method, summarizing the characteristics of the previous work, and advancing a new local color transfer method.2) According to the characteristics of the image, bringing the image Gaussian pyramid as an image region matching constraints and improving the accuracy of image region matching.3) Importing the idea of classified image database into the processes of region matching as guidance to improve the stability and reliability of the algorithm, while reducing the dependence on the results of image segmentation. The results show that the two methods are effective. Compared to existing methods, these methods employ the characteristic information of the image itself as much as possible, thus the algorithms reduce user's interaction effectively, and improve the accuracy and reliability of color transfer as well.