小波变换与纹理合成相结合的图像修复

目的 为了克服传统的图像修复算法在结构和纹理边界的错误修复,利用小波变换域的系数特征,探讨了一种基于小波变换与纹理合成相结合的修复算法。方法 算法先利用小波变换将待修复图像分解成具有不同分辨率的低频子图和高频子图,然后根据不同子图各自的特征分别进行修复。对代表图像结构信息的低频子图,采用FMM(fast marching method)算法进行修复;对代表图像纹理信息的高频子图,根据各子图中小波系数的特征,利用纹理合成方法进行修复。结果 分层、分类修复方法对边缘破损具有良好的修复效果,其峰值信噪比相比于传统算法提高了1~2 dB。结论 与相关算法相比,本文算法的综合修复能力较好,可以有效修复具有较强边缘和丰富纹理的破损图像,尤其对破损自然图像的修复,修复后图像质量得到较大提升,修复效果更符合人眼视觉效应。

Image inpainting based on combination of wavelet transform and texture synthesis

Objective An image inpainting algorithm based on combination of wavelet transform and texture synthesis is discussed to overcome the error repair of the boundary of structure and texture in traditional image inpainting algorithm. The discussed image inpainting algorithm utilizes characters of wavelet transform domain coefficients. Wavelet transform has been used as a good image representation analysis in addition to statistical properties. Multiresolution analysis of wavelet transform is helpful to predict coarse-to-fine image structure. In particular, texture and detailed patterns for natural images must be analyzed. Wavelet can treat these elements altogether. In view of the advantages of image decomposition algorithm, wavelet coefficient statistical properties, and visual effect of edge information of an image, we proposed an image inpainting algorithm based on combination of wavelet transform and texture synthesis. Method Our reconstruction modeling is based on classical image decomposition model. Some actions have been taken to improve reconstruction performance. An image can be seen as a combination of texture and structure. Thus, the image repair process should fully consider the texture and structural characteristics of an image. At first, the damaged image is decomposed into low-frequency sub-image and high-frequency sub-image with different resolutions via wavelet transformation. In cases where low-frequency component represents image structure, high-frequency component reflects edge changes of an image. Moreover, low-frequency component has a positional correspondence relationship with high-frequency component. Then, sub-images are reconstructed in accordance with their respective characteristics. The sub-image that reflects structural information of an image is reconstructed with fast multipole method, whereas the sub-image that reflects texture information of an image is filled in with texture synthesis based on the characteristics of wavelet coefficient in sub-images. We introduce edge factor in combination with the characters of the wavelet transform domain coefficients to update priority function in the process of reconstituting high-frequency sub-images. Finally, the recovered sub-images are reconstructed with wavelet. Result Simulation results show that this hierarchical classification method works well in edge damaged blocks. The power signal-to-noise ratio of the final result compared with the traditional algorithm has been improved by approximately 1 dB to 2 dB. The repair results are consistent with human visual perception. Conclusion Image decomposition model is a widely used image inpainting method. However, fuzzy and mismatching can be generated easily during the repair process. Therefore, when high-frequency component is repaired, changes in factor coefficients of high-frequency components must be introduced to enable the repair process be in accordance with edge direction. In such case, repairing image edge and improving matching block search are top priorities to reduce mismatch error. The proposed method can eliminate point defects in the repair process. Compared with the related algorithms, our algorithm holds good integrated performance. It can effectively repair damaged image with strong edges and rich texture, particularly for the loss scenarios and natural images, to improve image inpainting quality and to be consistent with human visual effects.