부산대학교 도서관

Deep learning has made significant advancements in single image super-resolution (SISR). However, the majority of research now in existence concentrate on generating networks that are more intricate and have a large number of layers, which causes the difficulties outlined above and reduces the effectiveness of the image formed. This work suggests an auto encoder embedded Caps Net that makes use of capsules. First, the auto encoder is used to retrieve pixel characteristics, greatly reducing noise in the raw data. Using a routing technique, the model generates multiple embeddings for every capsule, resulting in class capsules and high-level encoded capsules. Capsules of different sizes are handled by an attention module, which enables the model to concentrate on significant pixel regions. In order to produce capsule output and prevent low-frequency information during training, the dual residual route is utilised. In order to retain inherent information, the feature maps from each capsule are concatenated and sent to the relative weighting block. A real-valued tensor of each up-sampled pixel is represented by a channel created by the magnification layer. This method creates single images with super resolution using deep learning, which produces effective pixel quality and highly preserved prominent attributes. The result reveals that the proposed method attained the high PSNR and SSIM values when compared to existing method.