A Deep Convolutional Neural Network for Super Resolution via Soft-Attention Mechanism

カテゴリ: 部門大会

論文No: SS1-1

グループ名: 【C】2021年電気学会電子・情報・システム部門大会

発行日: 2021/09/08

タイトル(英語): A Deep Convolutional Neural Network for Super Resolution via Soft-Attention Mechanism

著者名: 張 伯聞（甲南大学）,田中 雅博（甲南大学）

著者名(英語): Bowen Zhang (Konan University),Masahiro Tanaka (Konan University)

キーワード: Deep Learning|Attention Mechanism|Convolutional Neural Network|Super Resolution

要約(日本語): Deep Convolutional Neural Networks(DCNNs) have achieved a state-of-the-art performance in computer vision tasks. But convolution operation can only process a local neighborhood at one time. This work constructs a DCNNs structure to adopt a soft-attention module for capturing the dependencies of spatial position information in each image, which is applied to a Super Resolution task. In addition, we use a few convolution filters with different kernel sizes in parallel for extracting the multi-scale features. In the experimental study, we compare the performance of our model with the traditional Bicubic interpolation method and the deep learning method SRCNN(3 layers DCNNs), where the Peak Signal-to-Noise Ratio index of our model is superior to the other two.

PDFファイルサイズ: 665 Kバイト