Image Inpainting Based On Scene Transform And Color Transfer

Digital image inpainting is the technique which provides a means for reconstruction of the missing or damaged parts of an image. According to the texture or color information provided, it can find optimal matching information for the target area. After that, it restores the image base on certain formula, making the restored image appear harmonious and natural. As a hotspot in current image processing, computer graphics and computer vision research fields, image inpainting is widely applied in image editing, virtual reality, video special effects manufacture etc.At present, most of image inpainting methods employ the information brought by target image itself, when it comes to larger ruined region. The target image always fails to offer sufficient information. As a result, it shows dissatisfied inpainting result in vision. This paper proposes an image inpainting algorithm based on scene transform and color transfer. At first, source images are obtained by searching for similar images in the image database. Then, the rough scene fill is performed on the target region by selecting suitable referenced content from source images. Finally, color transfer is implemented so that the restored region appears in harmony with the surrounding environment in the image. The experimental results show that the restored results improved greatly in vision quality due to the information introduced from the reference source images.Image quality evaluation is very important for analyzing the performance of image inpainting algorithms. A direct way to evaluate the effect of the reconstruction is to compare the difference in content between restored image and original image without damage. However, original images without damage are not available in most cases, which is also the origin of image inpainting. Moreover, the traditional image quality evaluations, such as PSNR, SSIM and so on, which all belong to a Full-reference evaluation model. Hence, they are not suitable for the evaluation of image restoration. Based on the above reasons, as for the impainting result of natural images after the substitution of the large area scene, this paper presents an assessment metric without reference. This metric makes a decision on whether the impainting vision results are good or not, by observing whether there existing a harmony relations between the filling area and the surrounding area. First, it will compute the color polymerization of pixels on the two sides of the impainting boundary, and divide the area into texture or flat part. Then, it would match the similarity between the two side connection area of the impainting boundary, and according to the varied importance, different areas will be endowed with different weights based on the human eye interests' theory. Finally, we obtain the assessment metric after weighting. Empirical results demonstrate that this assessment metric has satisfied correlations with the subjective assessment results, which can efficiently reflects the subjective quality of the image after impa(?)nting.