http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
- 中文 을 입력하시려면 zhongwen을 입력하시고 space를누르시면됩니다.
- 北京 을 입력하시려면 beijing을 입력하시고 space를 누르시면 됩니다.
RISS 인기검색어
- 검색키워드
- 저자 : Hosna, Asma-ull (검색결과 3 건)
-
Semantic segmentation of 3D medical image using guided genetic adversarial network
Hosna, Asma-ull Hankuk University of Foreign Studies. Graduate Sch 2018 국내석사
Semantic Segmentation is one of the core phenomena which can be dealt inside the verge of Artificial Intelligence World. In this approach, each pixel of an image is required to be labeled to yield the final segmented result. In this thesis, we proposed a novel method which is conducted following by state-of-the-art of Generative Adversarial Network (GAN) for Semantic Segmentation. To the best found knowledge, this is the first ever approach for Semantic Segmentation where the actual mainstream along with guided GAN is instigated. We have trained a Generator and Discriminator combining a deconvolution and traditional convolutional semantic segmentation network, respectively. Our work can be well-thought-out as a frontier for Semantic Segmentation as its adopting capability of the true nature of GAN. In this method, we process the original 3D medical image into sub-sample to pass through multiple image channels which work as input of GAN. This subsample, however, gradually generate the segmentation mask for the corresponding input image. The key insight of our work is that traditional GAN can only generate the fake image from some random noise based on input image without any guidance whereas our approach generates guided segmentation mask for corresponding input image instead of generating only the image replica. Our proposed methods were tested to perform segmentation from various image conditions (e.g. MRI, CT) of several human organs (e.g. Aortic Root, Left Atrium, Upper and Lower Cartilage of Knee, Retinal Vessel, Brain) where a significant amount of accuracy improvement has observed. Here, we compare our result of different GAN architecture with the traditional method. Such system, hence, will support many experiments to help better understanding of Humankind in the perspective of Artificial Visualization. We also intend to extend the platform to maximize the likelihood of the desired description for a complex and diverse scenario of various environments.
-
Hosna Asma ull 한국외국어대학교 대학원 2023 국내박사
Machine learning techniques for classification frequently presumptively use unordered class values. However, the class values do show a natural order in many practical applications. For instance, while learning how to rate the severity of a patient's sickness, estimate age with various image datasets. Ordinal classification often involves converting the class value into a numerical quantity, applying a regression learner to the transformed data, and then, in a post-processing step, transforming the output back into a discrete class value. Ordinal prediction is a halfway solution to classification and regression-based problems. For ordinal prediction of diverse regression-based problems, we proposed a noble Zig-Zag cascade in this thesis, based on deep learning. We also incorporate multi-scale CNN to enhance the performance of the suggested approaches. In order to stabilize the performance of the suggested approaches, we also concentrate on the feature of the order relationship among distinct classes. We suggest improvements in three separate areas for ordinal prediction in various regression-based problems in this thesis. First, we propose, first ever, direct ordinal prediction to estimate the severity level of Breast arterial calcifications (BACs) without prior segmentation of calcified vessel, unlike earlier methods. Other researcher marked pixel-to-pixel ground-truth (GT) for calcified vessel which is only can done by experience experts. As a consequence, the maintenance of the dataset is costly and time consuming. We customized ResNet50 to adjust for small object in bigger image for direct ordinal prediction of Mammogram. Second, we propose ensemble model to improve the performance of intra-class of RSNA bone dataest to estimate age. Here, we use Customized ResNet50 as deep model. However, unlike other traditional method, we make sub-group based on edge segmentation (without additional ground truth marking) and use as input of third channel. Afterwards, Customized ResNet50 estimate age individually. Finally, we, combine the predicted individual result and analysis the result. Third, we proposed a noble Zig-Zag cascade for ordinal prediction by emphasizing the order relationship among various classes. Unlike other cascade method, proposed Zig-Zag cascade avoid classifying neighboring classes as those have less distinctive feature to compare. This method classify extreme classes gradually and narrow-down the training and testing dataset. We use wide range of dataset from different fields such as Mammogram, Face dataset, Bone dataset, to justify the idea of proposed method.
-
연관 검색어 추천
이 검색어로 많이 본 자료
활용도 높은 자료
