Sida Zheng

Nighttime images suffer from exposure imbalance, leading to significant challenges in nighttime semantic segmentation. Given this, we propose a frequency-based dual attention network (FDANet) for nighttime semantic segmentation. To address the problem of targets hidden in the background caused by uneven exposure and low contrast that are difficult to distinguish, we propose a multi-spectral dual attention module to deeply explore hidden features from frequency information. In addition, to further utilize the relationship between frequency domain information and spatial domain features and guide the extraction of contextual features, spatial-frequency aggregation module is proposed to align and fuse features from different domains. Moreover, self-attention dilated convolution pyramid module is proposed to expand the effective receptive field and enhance the ability to identify multi-scale targets under uneven exposure. Extensive experiments on NightCity, ACDC, and BDD show that FDANet achieves an impressive balance between accuracy and speed. Specifically, FDANet achieves 53.76% MIoU, 49.47% MIoU, and 39.70% MIoU on NightCity, ACDC-Night, and BDD-Night, respectively. Notably, it attains these results with remarkable processing speeds of 111 FPS, 66 FPS, and 104 FPS for input resolutions of 512×1024, 1080×1920, and 720×1280, respectively.

Due to the excellent semantics extraction capabilities, deep learning methods have significantly progressed in salient object detection (SOD). However, these methods often require time-consuming pre-training and large training datasets with ground truth. To address these issues, by referring to the framework known as “deep image prior (DIP),” we propose a SOD method called deep label prior network (DLPNet), which consists of $\cal A$ -stream and $\cal B$ -stream. The $\cal A$ -stream includes two cascaded UNets and a simple CNNs module to extract the initial saliency map, while the $\cal B$ -stream contains only two cascaded UNets, which refines the extracted initial saliency map. Unlike most of the current deep learning methods, DLPNet views the SOD task as a conditional image generation problem, relying on only the internal prior of the input itself to generate the saliency map. Hence, our DLPNet does not require pre-training or large annotated / unannotated datasets. Furthermore, we propose a morphology operation scheme, which creates rich pseudo-labels for facilitating the updating of network weights. Extensive experiments demonstrate that our method outperforms state-of-the-art unsupervised techniques and is even comparable to state-of-the-art supervised and weakly supervised methods on different evaluation metrics.