Full metadata record
DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | 정제창 | - |
dc.contributor.author | Shuting WANG | - |
dc.date.accessioned | 2022-09-27T16:12:58Z | - |
dc.date.available | 2022-09-27T16:12:58Z | - |
dc.date.issued | 2022. 8 | - |
dc.identifier.uri | http://hanyang.dcollection.net/common/orgView/200000628137 | en_US |
dc.identifier.uri | https://repository.hanyang.ac.kr/handle/20.500.11754/174614 | - |
dc.description.abstract | With the improvement of computer hardware computing power and the continuous development of artificial intelligence technology, deep learning has achieved outstanding performance in the field of computer vision. At this stage, super-resolution (SR) based on deep learning mainly uses convolutional networks to recover texture details. However, the receptive field of such neural networks depends on the size of the filter and the number of convolutional layers. Increasing the values of these hyper parameters can increase the complexity of the model or even make it impossible to train. The long-distance dependencies of the convolutional feature extraction process tracking images can make the model quite complex. So convolutional feature extraction has the disadvantage of not being able to obtain long-range dependencies. The transformer, on the other hand, can model longer distance dependencies by using the self-attention mechanism. In this paper, we carry out the research of image SR reconstruction method based on deep learning, and the main research contents and contributions are as follows: (1) In order to better utilize the high-frequency information of the image, an attention-based image SR reconstruction method is proposed. The method introduces attention in the network. The method introduces the attention mechanism in the network, explores a lightweight transformer structure, and uses the self-attentive capability of the transformer to establish a long-range dependency for the network. (2) Proposed a bridging feature extraction structure based on convolution and transformer. The transformer structure is combined with CNNs to establish distal dependencies for the transformer by convolving the input fine local texture details, while the feature representation obtained by the transformer through global similarity modeling can provide distal information for the convolution. Thus, it assists the image SR network to establish distal dependencies and complete the extraction and recovery of high-frequency information from images. In this paper, we use Flickr2k as the training set and extract a part of it as the test set. Experimental results show that by comparing PSNR and SSIM with other SR algorithms, the transformer bridge-based SR n method proposed in this paper produces better SR images compared with the state-of-the-art methods, and our model consistently produces more distinct edges and shapes, effectively improving the visualization of the images. | - |
dc.publisher | 한양대학교 | - |
dc.title | Single Image Super-Resolution Based on Transformer Bridge | - |
dc.type | Theses | - |
dc.contributor.googleauthor | 왕서정 | - |
dc.contributor.alternativeauthor | 왕서정 | - |
dc.sector.campus | S | - |
dc.sector.daehak | 대학원 | - |
dc.sector.department | 융합전자공학과 | - |
dc.description.degree | Master | - |
dc.contributor.affiliation | 고속 인터페이스 | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.