Hierarchical attention network with progressive feature fusion for facial expression recognition.

Facial expression recognition (FER) in the wild is challenging due to the disturbing factors including pose variation, occlusions, and illumination variation. The attention mechanism can relieve these issues by enhancing expression-relevant information and suppressing expression-irrelevant information. However, most methods utilize the same attention mechanism on feature tensors with varying spatial and channel sizes across different network layers, disregarding the dynamically changing sizes of these tensors. To solve this issue, this paper proposes a hierarchical attention network with progressive feature fusion for FER. Specifically, first, to aggregate diverse complementary features, a diverse feature extraction module based on several feature aggregation blocks is designed to exploit both local context and global context features, both low-level and high-level features, as well as the gradient features that are robust to illumination variation. Second, to effectively fuse the above diverse features, a hierarchical attention module (HAM) is designed to progressively enhance discriminative features from key parts of the facial images and suppress task-irrelevant features from disturbing facial regions. Extensive experiments show that our model achieves the best performance among existing FER methods.

A Two-Stage Network for Zero-Shot Low-Illumination Image Restoration.

Due to the influence of poor lighting conditions and the limitations of existing imaging equipment, captured low-illumination images produce noise, artifacts, darkening, and other unpleasant visual problems. Such problems will have an adverse impact on the following high-level image understanding tasks. To overcome this, a two-stage network is proposed in this paper for better restoring low-illumination images. Specifically, instead of manipulating the raw input directly, our network first decomposes the low-illumination image into three different maps (i.e., reflectance, illumination, and feature) via a Decom-Net. During the decomposition process, only reflectance and illumination are further denoised to suppress the effect of noise, while the feature is preserved to reduce the loss of image details. Subsequently, the illumination is deeply adjusted via another well-designed subnetwork called Enhance-Net. Finally, the three restored maps are fused together to generate the final enhanced output. The entire proposed network is optimized in a zero-shot fashion using a newly introduced loss function. Experimental results demonstrate that the proposed network achieves better performance in terms of both objective evaluation and visual quality.

Lightweight Single Image Super-Resolution with Selective Channel Processing Network.

With the development of deep learning, considerable progress has been made in image restoration. Notably, many state-of-the-art single image super-resolution (SR) methods have been proposed. However, most of them contain many parameters, which leads to a significant amount of calculation consumption in the inference phase. To make current SR networks more lightweight and resource-friendly, we present a convolution neural network with the proposed selective channel processing strategy (SCPN). Specifically, the selective channel processing module (SCPM) is first designed to dynamically learn the significance of each channel in the feature map using a channel selection matrix in the training phase. Correspondingly, in the inference phase, only the essential channels indicated by the channel selection matrixes need to be further processed. By doing so, we can significantly reduce the parameters and the calculation consumption. Moreover, the differential channel attention (DCA) block is proposed, which takes into consideration the data distribution of the channels in feature maps to restore more high-frequency information. Extensive experiments are performed on the natural image super-resolution benchmarks (i.e., Set5, Set14, B100, Urban100, Manga109) and remote-sensing benchmarks (i.e., UCTest and RESISCTest), and our method achieves superior results to other state-of-the-art methods. Furthermore, our method keeps a slim size with fewer than 1 M parameters, which proves the superiority of our method. Owing to the proposed SCPM and DCA, our SCPN model achieves a better trade-off between calculation cost and performance in both general and remote-sensing SR applications, and our proposed method can be extended to other computer vision tasks for further research.