Issn 2091-5446 ilmiy axborotnoma научный вестник scientific journal
Download 1.19 Mb. Pdf ko'rish
|
ilovepdf merged
- Bu sahifa navigatsiya:
- Conclusions
ILMIY AXBOROTNOMA
INFORMATIKA 2021 - yil, 1 - son 112 From Table (1), we can notice that the accuracy of identifying UzSL based on gestures achieved very high results. It can be also seen that the accuracy slightly decreased when the number of training images are decreased. Moreover, it can be remarked that the accuracy achieved by applying SVM classifier is better than the one accomplished by applying the minimum distance and k-NN classifiers. Also, we note that, SIFT achieved excellent results. But, after applying LDA the feature length becomes only 210 features and then the classification process becomes more faster. Table (2) shows that the best accuracy achieved when the PeakThr=0.0, Psize=16 × 16 and Nangels is 8. When the PeakThr parameter increases, the number of features decreases and then the accuracy decreases. The highest accuracy achieved when Psize=16 × 16. When the Psize increases, then SIFT will consider as global features and decreases Psize will not extract more features which are necessary in identification process. Increasing the Nangels and Nbins will extract and collect features in different orientations and solve rotation problem. Conclusions In this paper, we have proposed a system for UzSL recognition based on gesture extracted from Uzbek sign images. We have used SIFT technique to extract these features. The SIFT is used as it extracts invariant features which are robust to rotation and occlusion. Then, LDA technique is used to solve dimensionality problem of the extracted feature vectors and to increase the separability between classes, thus increasing the accuracy for our system. In our proposed system, we have used three classifiers, SVM, k-NN, and minimum distance. The experimental results showed that our system has achieved an excellent accuracy around 98.9%. Also, the results proved that our approach is robust against any rotation and they achieved an identification rate of near to 99%. In case of image occlusion (about 60% of its size ), our approach has accomplished an accuracy (approximately 50%). In our future work, we are going to find a way to improve the results of our system in case of image occlusion and also increase the size of the dataset to check its scalability. Also, we will try to identify characters from video frames and then try to implement real time UzSL system. Download 1.19 Mb. Do'stlaringiz bilan baham: |
Ma'lumotlar bazasi mualliflik huquqi bilan himoyalangan ©fayllar.org 2024
ma'muriyatiga murojaat qiling
ma'muriyatiga murojaat qiling