Screening Texture Classification

There are diverse screening patterns in the screened images; each can be best removed by a specific kind of filters, so the lack of the screening information in the screened images restrains the descreening results and complicates this kind of problems. Textures can be viewed as complex visual patterns composed of entities, or sub-patterns, that have characteristic brightness, color, slope, size, etc. [44]. Generally speaking, four kinds of popular approaches nearly dominate all the researches in the texture analysis, i.e., structural, statistical, model-based, and transform methods.

Structural approaches [44] ~ [45] represent texture by well-defined primitives (micro-texture) and a hierarchy of spatial arrangements (macro-texture) of those primitives. As to model-based texture analysis [46] ~ [52], it uses fractal and stochastic models, and attempts to interpret an image texture by using generative image model and stochastic model, respectively. Finally, transform methods for texture analysis, such as Fourier [51], Gabor [52] ~ [53] and wavelet transforms [54]

~ [56] represent an image in a space whose co-ordinate system has a strong understanding that is closely connected with the characteristics of a texture, like scales or frequency.

Our proposed texture analysis scheme combines the advantages of structural methods with those of statistical ones. The screening patterns can be referred to as primitives in structural methods since the classification mechanism is a supervised one. In the descreening phase, smooth indices are used to determine whether the screening patterns could stand for the screenings in the original images. If so, the screening pattern extracted from the original image will be fed into our classification engine, GA-CNN. As soon as the type of screenings in the testing image is identified, the following processes will be much easier and the descreening performance will be

better.

3.3.1 Screening-Texture Patterns

Screening patterns can be viewed as one kind of textures that represents the similarity grouping in an image. It may be difficult to get this sort of similarity in the screened images accurately. It takes no effort, however, to observe the regularity of the same screened images by human perception. For example, Fig. 3.3.1_1 (a)-(c) contain different types of screening patterns, each of which has its own regularity or uniformity, called texture of its own. Fig. 3.3.1_1 (a) contains screenings with smaller granulations; while Fig. 3.3.1_1 (b) has bigger or squared screenings. For the image corrupted by a specific screening pattern, a proper filter should be used for descreening. The total number of screening patterns for all the screened images can not be known beforehand. How to determine the number of screening patterns depends on the desired functional performance. In our experiments for the descreening purpose, only two classes of screening patterns are classified. Our experiments have showed that the descreening performance after two-class screening classification is very satisfactory and acceptable to human perception. The two classes of screening patterns that are cropped manually from the screened images in our database are showed in Fig. 3.3.1_2 and Fig. 3.3.1_3, respectively. These patterns and manual classification results will be used for the training of the proposed GA-CNN texture classifier.

In the testing phase of the trained GA-CNN texture classifier, it is essential to identify the proper block(s) in the testing image for screening texture identification and classification. We shall propose a set of smooth indices calculated from an image block for determining whether the extracted block is qualified to be one of the screening patterns in the testing images. Smooth indices thus play a critical role in the screening classification. The derivation in more details for smooth indices will be

depicted in the next subsection.

(a) (b)

(c)

Fig. 3.3.1_1 Three example screened images. Images (a) – (c) contain different types of screening patterns.

Fig. 3.3.1_2 Screening Example 1

Fig. 3.3.1_3 Screening Example 2

3.3.2 Smooth Indices for Screening-Texture Block Detection

The standard deviations in statistical approaches were always used as the analytical tool for signal processing and image processing. The smooth index used in our approach can be also represented in terms of the standard deviation; it is obtained from the proportion of difference in the standard deviations. The standard deviation in an image represents the extent of difference in intensities of an image. Looking for an index which can best describe how smooth a block will be, we have made use of the proportion of difference in the standard deviations between our defined blocks, named the difference ratio of standard deviations.

The size of the screening pattern that we extracted from one of screened images is 64 by 64 pixels. We partition the extracted screening pattern into five parts (i.e., sub-screening patterns). Figure 3.3.2_1 illustrates this partition of the screening pattern. For each of them, a standard deviation value has to be calculated, and then the calculated value needs to be compared with that of the original screening pattern.

With regard to all of these five sub-screening patterns, five difference ratios can be acquired for a screening pattern in a screened image. In the same way, four difference ratios can also be obtained about the central part from which is partitioned off the screening pattern. None of these nine numerical values being larger than ten percents totally makes sure of the smoothness of the extracted screening pattern. This smooth index can be tuned higher if the original document image is not that smooth and the components of edges may be more dominant. Although the smooth index is relevant to the screened images, it is ranged from 10% to 20% without respect to the complexity of the screened images. As Fig. 3.3.2_1 shows, each of the five blocks has half of the size of the original screening pattern, 32 by 32 pixels. The higher the smooth index is, the coarser the extracted screening pattern is. A smaller smooth index implies the uniformity and regularity of the extracted screening pattern. The similarity among blocks can truly reveal the agreement to the screening pattern coming from the testing screened image.

Fig. 3.3.2_1 Partition in the screening pattern for calculating the smooth indices

3.4 Design of CNN Templates for Screening Texture