Low-light image enhancement is a low-level task that aims to improve the brightness and contrast of underexposed images. State-of-the-art low-light methods for image enhancement, benefiting from the development of deep learning, have made great progress. The majority of existing approaches, however, are based on convolutional neural networks, and just a few attempts have been made with Transformers which could perform admirably on high-level vision tasks by readily building long-range dependencies and global context connections. In this paper, we propose LLU-Swin, a hierarchical encoder-decoder architecture, based on Swin Transformer and U-Net for low-light image enhancement. LLU-Swin is composed of several Residual Recovery Transformer Modules (RRTM), each of which contains several improved Swin Transformer layers with a residual connection. We also show Dilated Local-enhanced Window Transformer Block (DLTB), which uses non-overlapping window-based self-attention to offer tremendous efficiency and employs Dilated Locally-enhanced Feed-Forward Network(D-LeFF) to enhance the local sensory ability. Benefits from these two designs, LLU-Swin has great power to capture both global and local contexts. Experimental results demonstrate the LLU-Swin outperforms the state-of-the-art methods in image quality and inference speed.

Co-salient object detection is a challenging task, which aims to segment the co-occurring salient objects in multiple images at the same time. To address this task, we propose an end-to-end Enhancement Location-Refinement Network (ELR-Net) to capture both salient and repetitive visual patterns from multiple images. For various scenarios, common objects in different images only have the same semantic information, so we first propose a deep co-salient method based on channel and spatial attention module (CASM), which combines the attention mechanism to enhance the common semantic information. Subsequently, we employ a Co-attention and Refinement Module (CARM) to capture the common attributes of co-salient objects by learning the features consensus representation from a group of images using our group affinity module (GAM) and then we develop a self-correlation module (SCM) to further fine-grained information on the co-salient regions. Specifically, SCM can maintain the feature independence upon semantic categories and further help our model to distinguish pixels with similar but different categories. Moreover, single image saliency maps (SISMs) are predicted to extract intra-saliency cues, and then a correlation fusion module (CFM) is employed to extract inter-saliency cues. The proposed ELR-Net is evaluated on three challenging benchmarks, i.e., CoSal2015, CoSOD3k, and CoCA, demonstrate that our ELR-Net outperforms 9 cutting-edge models and achieves state-of-the-art performance.

Salient object detection methods focus on detecting precise salient objects and acquiring clear boundaries, but many approaches fail to achieve both of them. To solve this problem, we present A Hybrid Salient Object Detection with Global Context Awareness. The proposed model consists primarily of Multi-level Feature Enhancement (MFE) and Top-layer Processing (TP). MFE module is a module that fuses high-level features, low-level features and corresponding global context information and it is used for foreground enhancement. TP module generates two kinds of features, one is the representative top-layer features, and the other is the global context information that acts on the corresponding MFE module. Comparing our method with other 12 state-of-the-art salient object detection methods with different evaluation metrics, the experimental results show that our method is superior to other methods.