Implementation of global and local thresholding algorithms in image segmentation of coloured prints

Transcription

1 Implementation of global and local thresholding algorithms in image segmentation of coloured prints Miha Lazar, Aleš Hladnik Chair of Information and Graphic Arts Technology, Department of Textiles, Faculty of Natural Sciences and Engineering, University of Ljubljana, Snežniška 5, SI-1000 Ljubljana; Abstract Segmentation of a digital image whose histogram does not allow for clear and easy differentiation between the features of interest and the background pixels, requires the use of an advanced thresholding algorithm. In our study two local, i.e. adaptive, thresholding procedures with various sizes of small picture regions (windows) were implemented on an image of a lab handsheet made of repulped coloured prints in order to accurately identify the location and the number of residual toner particles. One of the applied algorithms was based on the information maximization of the thresholded image while the other was a modified Niblack's method. Results of the study suggest that the latter method outperforms the former one when using a large window size. Information maximization algorithm typically fails to detect some lighter and coloured areas of objects of interest. Key words: image processing, image binarization, local thresholding 1. Introduction Segmentation of a digital image into two regions one containing objects of interest and the other consisting of background pixels is a crucial step in various image processing applications. In pulp and paper research, to evaluate performance of a printed paper recycling process (e.g. flotation deinking), it is necessary first to separate ink particles and other impurities from the unprinted areas, i.e. paper itself. This process, known as thresholding can be described as follows (Young et al., 2008): If a[m,n] < a[m,n] = object = 1 Else a[m,n] = background = 0 (1) where parameter normally denotes brightness threshold applied to the image a[m,n]. Here we are interested in dark objects on a light background. Digital images of a printed paper surface are typically acquired with a flat-bed scanner and are usually 8- bit monochrome images having 256 different grey levels ranging from 0 (black, particles and other impurities) to 255 (white, paper). For images that exhibit a pronounced bimodal grey level distribution one peak represents grey levels of ink and dirt particles, and the other of the unprinted background pixels setting the right threshold does not present a problem: it is located at the local minimum between both peaks (Antoine et al., 2001). Often, however, difference between the ink peak and the paper peak is not well-defined and the image histogram is characterized by an unimodal, rather than a bimodal curve shape. This happens, for instance, when there is a low contrast between an ink (e.g. yellow) and the paper. Here instead of simply setting one global threshold for the total image, a more sophisticated algorithm has to be implemented. One possibility is the use of a local (adaptive, dynamic) thresholding method where the threshold changes dynamically over the image, i.e. its values are computed locally for small region windows. Objective of our research was to evaluate applicability of the two more advanced algorithms for digital image segmentation: an approach based on the digital image information maximization and an adaptive, locally-based thresholding technique, as described below. 2. Methods In our study we investigated a 2.1 x 2.1 cm 32-bit RGB colour picture (Figure 1, left) of a laboratory handsheet made of repulped prints scanned at 300 dpi. Paper was originally printed on a digital offset 1

2 printing machine Indigo E-Print at 600V printing drum voltage so the handsheet contained coloured ink (toner) particles of various sizes and shapes. Relatively large areas of the residual particles are typical of this printing technique; their elimination in paper recycling process using conventional flotation deinking procedure has proven to be an extremely difficult task (Hladnik et al., 2006). To compare various thresholding algorithms implemented in our study and to monitor the effect of image type, we examined another digital image that was produced by scanning part of a black-and-white text document using uneven light illumination (Figure 1, right). Our goal in this example was to obtain a well thresholded, readable text inspite of the present top-right to bottom-left decrease in background brightness. Figure 1. Images examined in the study: lab handsheet made of repulped print (left) and text document with illumination gradient (right). Two thresholding techniques for image segmentation were applied. First one is the so called information maximization (IM) algorithm (Kapur et al., 1985) where the threshold t max maximizes the information of the thresholded image according to the formula: t max arg Max H0( t) H1( t) tg N (2) where H 0 and H 1 are defined as follows: P P H ( o t) and N 1 Pi Pi H 1( t) log (3) it 11 P ( t) 1 P ( t) t i i log i0 Psum ( t) Psum ( t) P i is related to the probability distribution of the relative frequencies P as described in the original article (Kapur et al., 1985). This procedure can also be used for determination of a threshold within a smaller picture region window (local IM). Here a window of a chosen size moves from pixel to pixel in the image so that for each of them an optimum threshold value for its nearest neighbourhood can be obtained. The second applied method was a modified local, i.e. adaptive thresholding algorithm that was originally developed by Niblack (Niblack, 1986). He suggested to compute the threshold from the data on local mean and local standard deviation over a certain region n x n window size (WS) around each pixel location. The local threshold T at a pixel (i, is calculated as: T(i, = m(i, + k v(i, (4) 2 sum sum

3 where m(i, and v(i, are the local sample mean and variance, respectively. Parameter k controls the effect of standard deviation due to objects features and while it can be a constant value for segmentation of, for example, document images, for grey level images having a lot of variations in grey levels it should vary according to their grey level distributions. Therefore Rais (Rais et al., 2004) proposed the following formula for calculating the value of the weight k: k = 0.3 ( m g( i, g( i, m l ( i, l ( i, (5) max(( m g( i, g( i,, m l ( i, l ( i, ) where σ g (i, is the global standard deviation of the total image and σ l (i, is the local standard deviation calculated for each window. The above described algorithms were implemented on a personal computer using the MATLAB software package. Description of our work is presented in Figure 2. First, the original RGB image (Figure 1, left) was converted into an 8-bit greyscale (GS) one according to the following weighted formula: Grey = Red Green Blue (6) This GS image was binarized using both global IM as well as local Rais' thresholding (RT) algorithm. In the latter one, we experimented with various WSs ranging from 31 to 251 pixels. Alternatively, we performed CMYK split of the original image to produce separate C, M, Y and K 8-bit images, on which both thresholding techniques as with the GS image were applied. The four acquired thresholded C, M, Y and K images were in the end combined into a composite CMYK binary image. Finally, as it is possible to vary WS also in the case of the IM algorithm, we tested a number of WSs on GS as well as on CMYK-split images (dotted rectangles in Figure 2) to see how altering this parameter affects thresholding efficiency and, consequently, detection of coloured pixels found in the original image. The initial text document picture (Figure 1, right) was already acquired as an 8-bit GS image, so here only left part of the scheme was performed, i.e. without CMYK split images. 3

4 Grayscale (GS) conversion Original image CMYK split GS image Cyan Magenta Yellow Black Information maximization (IM) Rais' thresholding (RT) IM RT IM global RT WS1 RT WS2... IM RT WS1 RT WS2... global IM WS1 IM WS2... IM WS1 IM WS2 IM local IM local Figure 2. Scheme of experimental work. 4

5 3. Results and discussion RT and IM thresholded, i.e. binary, images of the text document image using various WSs are shown in Figures 3 and 4, respectively, while the corresponding results for the repulped print handsheet image are displayed in Figures 5 to 8. Several researchers reported that the local thresholding techniques were particularly successful in binarizing document images (Rais et al., 2004). Our experiments with the text image (Figure 1, right) confirmed these claims. When RT algorithm with a small WS e.g. 31 pixels was used (Figure 3), uneven background illumination was removed to a large extent. With increasingly bigger WSs, however, more and more letters in the bottom-left part of the image merged with the dark background present in the original picture and were therefore unreadable. With the other applied algorithm IM the observed effect of WS was quite the opposite (Figure 4): here best results were obtained with larger WSs. As demonstrated in the right-most picture, global IM thresholding proved ineffective in properly segmenting the image. Figure 3. Text document image results: binary images obtained with RT algorithm using various WS: 31 (left), 141(middle) and 251 (right) pixels. Figure 4. Text document image results: binary images obtained with IM algorithm using various WS: 3 (left), 25, 57 pixels and IM global (right). Interpretation of results obtained with implementing the two above mentioned thresholding techniques on the repulped print handsheet picture (Figure 1, left) was, as could be expected, not as straightforward as with the text document picture. Partial overlapping of toner particles of various colours, presence of small objects belonging to paper, i.e. background, and not to particles of our interest; these and other peculiarities made successful segmentation of the original image more difficult and also called for final human judgement regarding the applicability of the particular algorithm. RT algorithm applied on GS image (Figure 5) and using small WS (e.g. 31 pixels) generated many artifacts, i.e. particles not present in the original picture. Increasing the WS lead to progressively better results and with WS = 251, i.e, when the window size equaled the image size (250 pixels), the detection of the majority of features of interest was very good. In other words, although RT is a local thresholding method, smaller WSs did not produce satisfactory results. RT is an extremely locally sensitive algorithm, so it detects even the slightest differences in GS values, i.e. intensity variations in digital images. 5

6 When using IM thresholding technique on GS image (Figure 6), smallest WS again resulted in generating spurious particle pixels while with larger WS the situation improved siginficantly in this respect, although the total number of successfully recognized and binarized object pixels was too low. On the other hand, with small WS (e.g. 14 or 25 pixels) the algorithm was successful in recognizing certain coloured residual particles (e.g. yellow) that remained undetected when using bigger WS. It therefore seems that certain trade-off in terms of WS is necessary, i.e. WS optimum should be somewhere between 14 and 57 pixels. Global IM (right-most image) exhibits further too pronounced reduction in the number of recognized toner particles. As depicted in Figure 2, the original picture was also split into four separate C, M, Y and K channels and thresholded separately in an attempt to properly binarize and detect coloured image areas, i.e. toner particles. After performing the RT or IM thresholding, the resulting images were again combined into a composite CMYK image. With RT thresholding (Figure 7) detection of coloured (yellow) and brighter particles was, as expected, very good. Excellent performance was again demonstrated when choosing the biggest possible WS, i.e. 251 pixels (right-most image). Similar discussion applies as in the case of combination GS image-rt procedure (see above). Performance of IM thresholding on CMYK split image (Figure 8) was similar, i.e. less than optimum, to that on GS image (Figure 6). The difference is that in the case of CMYK separation of channels the resulting binary images contained, at least to some degree, pixels of yellow toner particles that were present in the original picture. It can be summarized that when comparing the original, repulped print handsheet image with the thresholded, binary images, the most complete and accurate detection of coloured features toner particles provides RT technique with a very large WS (251 pixels) applied either on the CMYK split or GS image. IM algorithm is less appropriate as it fails to detect some lighter and coloured areas of objects of interest, although its performance on CMYK split image is still acceptable. Figure 5. Repulped print handsheet GS image results: binary images obtained with RT algorithm using various WS: 31 (left), 141 (middle) and 251 (right) pixels Figure 6. Repulped print handsheet GS image results: binary images obtained with IM algorithm using various WS: 3 (left), 25, 57 pixels and IM global (right) 6

7 Figure 7. Repulped print handsheet CMYK split image results: binary images obtained with RT algorithm using various WS: 31 (left), 141 (middle) and 251 (right) pixels Figure 8. Repulped print handsheet CMYK split image results: binary images obtained with IM algorithm using various WS: 3 (left), 25, 57 pixels and IM global (right) 4. Conclusions It is a well-known fact that the same thresholding method can be successfully implemented in one application, but may not work well for another one. As the results of our study indicate, adaptive, i.e. local thresholding techniques proved to be a promising way to accurately segment a digital image such as printed paper prints into two regions: one containing toner or ink particles together with other impurities and the other consisting of background paper pixels. In the future we will, apart from testing and studying other thresholding procedures, focus on further optimization and automation of the presented methods, for instance in terms of speed so that a larger picture area can be analyzed in a shorter time period. Literature Antoine C., Lloyd M.D., Antoine J.: A robust thresholding algorithm for halftone dots. Journal of Pulp and Paper Science, Vol. 27, No. 8, P Hladnik A., Bolanča-Mirković I., Bolanča Z.: Characteristics of recycled fibres from digital prints using solid and liquid toner. Prooceedings of 2nd International Symposium on Novelties in Graphics, Ljubljana, Slovenia, June Kapur J.N., Sahoo P.K., Wong A.K.C.: A new method for grey level picture thresholding using the entropy of the histogram. Computer Vision, Graphics and Image Processing, Vol. 4, P Niblack W.: An introduction to digital image processing. Englewood Cliffs, N.J.: Prentice Hall, P Rais N.B., Hanif M.S., Taj I.A.: Adaptive thresholding technique for document image analysis. INMIC P Young I.T., Gerbrands J.J., van Vliet L.J: Image processing fundamentals Segmentation: Thresholding. WEB: 7

8 Acknowledgement The authors would like to thank Prof. Dr. Zdenka Bolanča from the Faculty of Graphic Arts, University of Zagreb, for providing digital sample images that were used in the study. 8