Local texture-based color transfer and colorization

Abstract

International audienceThis paper targets two related color manipulation problems: Color transfer for modifying an image's colors and colorization for adding colors to a grayscale image. Automatic methods for these two applications propose to modify the input image using a reference that contains the desired colors. Previous approaches usually do not target both applications and suffer from two main limitations: possible misleading associations between input and reference regions and poor spatial coherence around image structures. In this paper, we propose a unified framework that uses the textural content of the images to guide the color transfer and colorization. Our method introduces an edge-aware texture descriptor based on region covariance, allowing for local color transformations. We show that our approach is able to produce results comparable or better than state-of-the-art methods in both applications

Automatic Example-Based Image Colorization Using Location-Aware Cross-Scale Matching

“…Fig. 5 illustrates the performance against several other popular example-based colourisation methods [14], [17], [18]. It is easy to find numerous matching errors in the results generated by these methods, e.g., the green colour is mapped above the blue sky in the first example, and the green colour is mismatched to the body of the pyramid while the blue colour is mapped to the grass, which is seldom found in the reference image in the second example.…”

“…To circumvent this, Pierre et al [17] proposed an improved method with a regularisation involving both luminance and chrominance information, which helps to better preserve edge structures in the colourised images. To improve texture matching, especially near edges, Arbelot et al [18] developed a colour transfer and colourisation method that utilises spatial coherence around image structure by adopting an edge-aware texture descriptor based on region covariance, although local matching is still performed independently. A locality consistent sparse representation learning method is proposed by [19].…”

Texture-Guided Multisensor Superresolution for Remotely Sensed Images

“…TGMS uses statistical texture descriptors presented in [36] based on region covariance [37,38] because of its efficient and compact way of encoding local structure and texture information via first-and second-order statistics in local regions.…”

“…Multiscale gradient descent [36] is performed on the upsampled-LR, down-up-sampled-HR, and texture-descriptor images to create their edge-aware versions. Here, "down-up-sampled" means a process composed of low-pass filtering, downsampling, and upsampling to generate a blurred version of the HR image, and "edge" refers to boundaries of objects recognizable in the LR image.…”

“…If the GSD ratio between the LR-HR images is large, some pixel values may not be replaced by those of correct objects because complex transitions between different objects are smoothed out (e.g., the water's edge in the color image of Figure 2b). To overcome this issue, the multiscale gradient descent is effective by iteratively performing the gradient descent while gradually blurring the HR guidance image at a larger scale [36]. In Figure 2c, we can see that the complex water's edge is aware in the color image.…”