Real-world time series analysis, such as healthcare, autonomous driving, and solar energy, faces unique challenges arising from the scarcity of labeled data, highlighting the need for effective semi-supervised learning methods. While the Virtual Adversarial Training (VAT) method has shown promising performance in leveraging unlabeled data for smoother predictive distributions, straightforward extensions of VAT often fall short on time series tasks as they neglect the temporal structure of the data in the adversarial perturbation. In this paper, we propose the framework of functional Virtual Adversarial Training (f-VAT) that can incorporate the functional structure of the data into perturbations. By theoretically establishing a duality between the perturbation norm and the functional model sensitivity, we propose to use an appropriate Sobolev ($H^{-s}$) norm to generate structured functional adversarial perturbations for semi-supervised time series classification. Our proposed f-VAT method outperforms recent methods and achieves superior performance in extensive semi-supervised time series classification tasks (e.g., up to $ \approx 9$ % performance improvement). We also provide additional visualization studies to offer further insights into the superiority of f-VAT.

Mural restoration holds significant importance for various reasons from cultural, historical, and artistic aspects. Although significant progress has been made in natural image restoration in recent years, mural image restoration remains a challenging problem due to its differences in structure, color, and style from natural images. Motivated by the structural and color distribution characteristics of mural images, we propose a novel two-stage edge-guided mural in-painting model that perceives global and local information through a Transformer-Multi-scale CNN block (TMC) and learns informative color cues by introducing an extra YUV color space. Extensive experiments on Mural and CelebA-HQ datasets show that our proposed approach outperforms the other state-of-the-art baselines in both qualitative and quantitative evaluations.

Abstract Low‐light images are captured in environments with minimal lighting, such as nighttime or underwater conditions. These images often suffer from issues like low brightness, poor contrast, lack of detail, and overall darkness, significantly impairing human visual perception and subsequent high‐level visual tasks. Enhancing low‐light images holds great practical significance. Among the various existing methods for Low‐Light Image Enhancement (LLIE), those based on the Retinex theory have gained significant attention. However, despite considerable efforts in prior research, the challenge of Retinex decomposition remains unresolved. In this study, an LLIE network based on the Retinex theory is proposed, which addresses these challenges by integrating attention mechanisms and a U‐Net‐like architecture. The proposed model comprises three modules: the Decomposition module (DECM), the Reflectance Recovery module (REFM), and the Illumination Enhancement module (ILEM). Its objective is to decompose low‐light images based on the Retinex theory and enhance the decomposed reflectance and illumination maps using attention mechanisms and a U‐Net‐like architecture. We conducted extensive experiments on several widely used public datasets. The qualitative results demonstrate that the approach produces enhanced images with superior visual quality compared to the existing methods on all test datasets, especially for some extremely dark images. Furthermore, the quantitative evaluation results based on metrics PSNR, SSIM, LPIPS, BRISQUE, and MUSIQ show the proposed model achieves superior performance, with PSNR and BRISQUE significantly outperforming the baseline approaches, where (PSNR, mean BRISQUE) values of the proposed method and the second best results are (17.14, 17.72) and (16.44, 19.65). Additionally, further experimental results such as ablation studies indicate the effectiveness of the proposed model.

Missing data imputation for location-based sensor data has attracted much attention in recent years. The state-of-the-art imputation methods based on graph neural networks have a priori assumption that the spatial correlations between sensor locations are static. However, real-world data sets often exhibit dynamic spatial correlations. This paper proposes a novel approach to capturing the dynamics of spatial correlations between geographical locations as a composition of the constant, long-term trends and periodic patterns. To this end, we design a new method called Dynamic Adjacency Matrix Representation (DAMR) that extracts various dynamic patterns of spatial correlations and represents them as adjacency matrices. The adjacency matrices are then aggregated and fed into a well-designed graph representation learning layer for predicting the missing values. Through extensive experiments on six real-world data sets, we demonstrate that DAMR reduces the MAE by up to 19.4% compared with the state-of-the-art methods for the missing value imputation task

It is important yet challenging to perform accurate and interpretable time series forecasting. Though deep learning methods can boost forecasting accuracy, they often sacrifice interpretability. In this paper, we present a new scheme of series saliency to boost both accuracy and interpretability. By extracting series images from sliding windows of the time series, we design series saliency as a mixup strategy with a learnable mask between the series images and their perturbed versions. Series saliency is model agnostic and performs as an adaptive data augmentation method for training deep models. Moreover, by slightly changing the objective, we optimize series saliency to find a mask for interpretable forecasting in both feature and time dimensions. Experimental results on several real datasets demonstrate that series saliency is effective to produce accurate time-series forecasting results as well as generate temporal interpretations.