A Wavelet-Based Asymmetric Convolution Network for Single Image Super-Resolution

Abstract

Recently, deep convolutional neural networks (CNNs) have been widely explored in single image super-resolution(SISR) and obtained remarkable performance. However, most of the existing CNNbased SISR methods tend to produce over-smoothed outputs and miss some textural details. To address these issues, we propose a wavelet-based asymmetric convolution network (WACN). Different from conventional CNN methods that directly infer HR images, our approach firstly learns to predict the LR's corresponding series of HR's wavelet coefficients before reconstructing HR images from them. This helps to capture more structural information in images to preserve texture information and avoid artifacts. To enhance the ability of feature extraction, we propose an asymmetric convolution block (ACB) structure to form a very deep network. In the training phase, ACB can provide different receptive fields to enrich feature information. In the inference phase, ACB's asymmetric convolution kernel can be equivalently fused into the standard square-kernel layers, such that no extra computational burdens are introduced in the inference phase. Furthermore, we propose a variance-based channel attention (VCA) mechanism to adaptively rescale channel-wise features by considering interdependencies among channels. Extensive experimental results demonstrate the superiority of the proposed WACN in comparison with the state-ofthe-art methods.

Asymmetric cost aggregation network for efficient stereo matching

“…\times \! k$ convolution to improve the network performance in the tasks of semantic segmentation [49] and image super‐resolution [47, 50].…”

“…Recently, some works follow the practice in ref. [46] and employ the skeleton-strengthened k×k convolution to improve the network performance in the tasks of semantic segmentation [49] and image super-resolution [47,50].…”

A Novel Levy Walk-Based Framework for Scheduling Power-Intensive Mobile Edge Computing Tasks

“…CNNs can utilize this multi-resolution decomposition property of wavelets by using convolutions to learn wavelet filters at each level [108]- [110]. The output of each level becomes the input to the next, with the filters extracting more detailed features at higher levels after the removal of coarse information.…”

“…The output of each level becomes the input to the next, with the filters extracting more detailed features at higher levels after the removal of coarse information. This convolutional learning of adapted wavelet bases enables CNNs to hierarchically capture patterns across different scales for improved data representation [110].…”

Image-based classification of healthy and diseased bell pepper using the transfer learning method

“…Certain researchers also use rescaling to consider the interdependencies among the channels [12]. This is a challenging procedure that requires a delicate balance of sharpness, efficiency, speed, and smoothness.…”