Handwritten Word Spotting in Old Manuscript Images using Shape ...

More documents

Recommendations

Info

Figure 18: Distribution of the observations in the clusters using basic features. The ideal solution in the clustering process is to obtain a 100% in completeness and homogeneity. In our case we have not obtain an ideal solution, then, we have to choose a measure which is a trade of between both measures. In figure 21 we observe two plots for each experiment, β = 0 means that the plot is measuring homogeneity and β = 1 means that the plot is measuring completeness. For each experiment we observe that with small number of clusters the homogeneity is small and the completeness is good. By increasing the number of clusters the homogeneity increases and the completeness decreases. The best number of cluster for each experiment is when both plots cross. For example, the best number of clusters for the BSM features is 15. The experiment for the retrieval process evaluates its accuracy. We have done several experiments using different combinations of basic features, the subset of the ground truth and the BSM features (Fig. 22). We observe that the worst results are obtained when we use the ground truth with all the basic features. Using the BSM features we have obtained the best results, followed by the experiment using the basic features height and width. Using all the basic features we have obtained worst results. In the last experiment we evaluate the performance in terms of scalability (an increasing number of documents and classes) and the descriptor. We observe, using the same descriptor and different number of documents and classes, that the accuracy is better with less number of classes. We also observe that using the BSM descriptor, it is a better descriptor and more accurate. The performance improves, even using the bigger ground truth with respect the best result of the smaller ground truth. 26
Figure 19: Distribution of the observations in the clusters using BSM features. In conclusion the performance is more sensitivity to the accuracy (descriptive power) of the descriptor. With the same descriptor the more is the number of classes, the higher is the confusion, so the performance decreases. Pseudo-Structural descriptor organized in a Hash Structure Our second approach is evaluated by using precision-recall curves. These experiments are done by tuning two parameters: mask size and the threshold used to decide if an observation is member of a class, or not. To obtain Loci features we have used different masks to obtain the number of intersections for each pixel in all directions (Fig. 23). There are two options in the feature extraction step. The first one is using the background pixels as reference to obtain the feature vector. The second experiment uses foreground pixels as reference. In figure 24b we observe the results using background pixels as reference, for different mask sizes and varying the threshold with the following values: 25, 50, 100, 200, 300, 400, 500 and 600. We observe that when we increase the size of the mask, the results are better. But, when the size of the mask is 80 pixels, or more, the results are very similar. We do not get more information because the mean of the height of the words is 80 pixels, then, increasing the size of the masks. The use of the foreground as reference pixels gives a similar performance (Fig. 24a). But if we compare the results using foreground and background as pixel reference (Fig. 25), we observe that 27
Page 1: MASTER IN COMPUTER VISION AND ARTIF
Page 4 and 5: a step forward towards shortening t
Page 6 and 7: a classification of the training se
Page 8 and 9: (a) 1617: index of volume 69 (b) 17
Page 10 and 11: The models can then be used to retr
Page 12 and 13: Figure 5: We present two approaches
Page 14 and 15: 6.1.1 Binarization The binarization
Page 16 and 17: partial Gaussian derivatives along
Page 18 and 19: 7. Pixel-based descriptors organize
Page 20 and 21: layer uses an automatic method. It
Page 22 and 23: Table 1: Intervals for each directi
Page 24 and 25: The segmentation process experiment
Page 26 and 27: Table 2: Pre-processing results. Th
Page 30 and 31: Figure 20: Comparative using differ
Page 32 and 33: Figure 22: Classification process u
Page 34 and 35: use foreground pixels, because the
Page 36 and 37: [15] G. Nagy. Twenty years of docum
Page 38: Figure 27: SOM using characteristic

Handwritten Word Spotting in Old Manuscript Images using Shape ...

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?