Texture detection in image processing

#An embodiment relates to the field of image processing and, in particular, the detection of the texture of a digital image.

#The image texture can be regarded as a set of metrics quantifying the arrangement of pixel intensities in a digital image or a selected region of an image. Texture detection is an important issue in image processing. For example, digital images, still pictures as well as videos, exhibit some noise that has to be reduced by appropriately designed algorithms. The noise filtering, however, has to be adjusted according to the detected texture in order to avoid blurring of the image. If a selected region of an image is homogenous (flat), it can be heavily filtered since pixel variations are basically caused by noise. If, on the other hand, the selected region of an image is highly textured, pixel variations are mainly caused by the texture, and filtering should be performed lightly and with caution. Basically, only small differences from pixel-to-pixel shall be noise filtered.

#In the art, texture detection is performed in the luminance domain and is based on high pass/low pass filtering. Conventional methods of texture detection are known as Sobel, Prewitt, or Canny maps. For example, the texture can be obtained from the Spatial Noise Level and the maximum of the absolute differences between a particular current pixel and its neighbors (see, for example, A. Bosco and M. Mancuso, “Adaptive Filtering for Image Denoising”, in IEEE Proceedings of ICCE2001, pages 208-209, 2001, which is incorporated by reference). However, despite the recent engineering progress, texture detection and noise reduction based on the same still are not sufficiently reliable and satisfying, in particular, in noisy images. The conventional methods not only are sensitive to noise, but also need appropriate fine tuning of thresholds in order to avoid a misclassification of pixels.

#Features and advantages of one or more embodiments will be described with reference to the drawings. In the description, reference is made to the accompanying figures that are meant to illustrate one or more embodiments. It is understood that such embodiments are not intended to be limiting.

#FIG. 1 illustrates a 9×9 processing kernel including a 3×3 sub-mask containing real Bayer samples and three 3×3 interpolated sub-masks used according to an example of a method for texture detection.

#FIG. 2 illustrates an example of the interpolation process used for texture detection based on the 9×9 processing kernel shown in FIG. 1.

#As it is shown in FIG. 1, according to an embodiment, a processing kernel having a size of 9×9 pixels is employed for analyzing a CFA raw data image. It is to be understood that an embodiment is not restricted to a size of 9×9 pixels and, particularly, a larger size can be chosen. In the CFA image each pixel location is sensitive to one color component, which is achieved by providing the appropriate filter for each of the pixels. Besides Red, Green, and Blue there are GreenRed representing a Green pixel in a row of Red pixels and GreenBlue representing a green pixel in a row of Blue pixels. An input CFA image is scanned pixel-wise by the moving processing kernel. For each position of the moving kernel (mask) a texture measure or a flatness measure and a texture or flatness degree are assigned to the central pixel. The mask is centered on that central pixel that can be a Red, GreenRed, GreenBlue, or Blue one. Note, that the degree of texture and the degree of flatness that are complementary to each other and can be alternatively measured and used in an embodiment. In fact, if no flatness is defined by a value of zero and 1 denotes maximum flatness, the degree of texture is given be 1 minus the degree of flatness.

#Once a texture/flatness degree is assigned to the central pixel, the mask moves to the next adjacent pixel or, alternatively, is moved by more than one pixel depending on the concrete realization. A 3×3 sub-mask centered on the central pixel contains real Bayer samples sensed by photodiodes covered by receptive color filters. Three other 3×3 sub-masks are obtained by interpolation over the same grid-locations where the real samples are available. An example for the interpolation is given in the following. Suppose at a particular time the moving kernel is centered on a GreenRed pixel, for example (see FIG. 1), i.e., the central pixel is a GreenRed one. A 3×3 sub-mask centered on the central pixel contains real sensed Bayer GreenRed samples indicated by the circles. However, in order to increase precision of the texture detection, especially along tiny edges with a width of about 1 pixel, the information from the other CFA channels, Red, Blue and GreenBlue, in the present example, is also taken into account. In order to exploit the information from the other CFA channels, pixels adjacent to the 3×3 sub-mask containing the real sensed Bayer GreenRed samples have to be considered.

#According to an example, the respective missing Bayer information at each pixel location where real sensed Bayer samples exist is interpolated. This means that the missing Bayer information (Red, Blue and GreenBlue) is interpolated at the pixels of real sensed Bayer GreenRed indicated by the circles in FIG. 1. This interpolation mitigates the problem of sample mismatch in the Bayer grid and, in particular, allows for a better detection of tiny lines of widths of about a pixel.