Transformers, renowned for their powerful feature extraction capabilities, have played an increasingly prominent role in various vision tasks. Especially, recent advancements present transformer with hierarchical structures such as Dilated Neighborhood Attention Transformer (DiNAT), demonstrating outstanding ability to efficiently capture both global and local features. However, transformers’ application in edge detection has not been fully exploited. In this paper, we propose EdgeNAT, a one-stage transformer-based edge detector with DiNAT as the encoder, capable of extracting object boundaries and meaningful edges both accurately and efficiently. On the one hand, EdgeNAT captures global contextual information and detailed local cues with DiNAT, on the other hand, it enhances feature representation with a novel SCAF-MLA decoder by utilizing both inter-spatial and inter-channel relationships of feature maps. Extensive experiments on multiple datasets show that our method achieves state-of-the-art performance on both RGB and depth images. Notably, on the widely used BSDS500 dataset, our L model achieves impressive performances, with ODS F-measure and OIS F-measure of 86.0%, 87.6% for multi-scale input,and 84.9%, and 86.3% for single-scale input, surpassing the current state-of-the-art EDTER by 1.2%, 1.1%, 1.7%, and 1.6%, respectively. Moreover, as for throughput, our approach runs at 20.87 FPS on RTX 4090 GPU with single-scale input. Code: https://github.com/jhjie/EdgeNAT.

Chinese spelling correction (CSC) is a crucial task in natural language processing, aiming to detect and correct spelling errors in Chinese text. The improved performance of Chinese spelling errors correction algorithms can enhance the efficiency and accuracy of upstream and downstream Chinese natural language processing tasks, such as OCR, ASR, and translation.However, current methods based on neural networks are mostly limited to either using only contextual information to correct misspelled words or failing to fully utilize glyph and pinyin information. Therefore, we propose a multimodal approach to address the above issues. Specifically, a three-tower multimodal structure is used to extract glyph, pinyin, and semantic information, and a decoder composing of an error probability prediction network and a transformer network is employed to achieve cross-modal information interaction. Besides, an additional training task is used to achieve cross-modal information alignment. Experiments demonstrate that proposed network outperform most existing motheds.

Taxi Origin-Destination (OD) demand prediction plays an important role in realizing intelligent cities. An accurate OD forecasting for taxi orders helps achieve efficient taxi dispatch, which reduces commuting costs for both drivers and passengers. Recently, graph convolutional networks (GCNs) have been widely used to model an OD Matrix as spatial-temporal graphs and reach remarkable performance. However, the conventional CNNs and GCNs always have fixed receptive fields, as the networks only aggregate nodes’ information among the k-th order neighborhood. These methods bring about tremendous computational complexity and parameters by introducing incremental modules and a multi-stream fusion strategy. In the case of taxi OD demand, the origin and destination are often far apart. To address these issues, we propose a novel clustering shift graph convolutional network (CSGCN). Specifically, we employ lightweight point-wise convolutions to extract the features, and a Shift operation is proposed to make the features circulate among every node which expands the receptive fields to the whole graph. Extensive experiments on two datasets demonstrate the superiority of our CSGCN over other GCN-based methods.

Traffic forecasting is one of the most fundamental components in many applications from urban computing to intelligent transportation. Recently, graph convolutional networks (GCNs), which model traffic data as spatiotemporal graph, have attracted lots of attention. However, existing GCNs mainly focus on pre-defined graph structure which is fixed over the entire network. These methods of traffic forecasting can not capture the complex spatial correlations especially for the higher level features. To address these problems, we propose a novel Multi-Adaptive Graph Convolutional Network (MAGCN) for traffic forecasting in this work. Our model can dynamically learn the topology of the graph through multi-range GCNs in an end-to-end manner. This data-driven method makes the construction of graph more flexible and increases the generality of model to adapt to various data samples. Moreover, in the proposed framework, we design a novel Differential Temporal Graph Convolutional Network (DTGCN) to capture the periodic and immediate temporal correlations of traffic data, which is integrated in MAGCN to effectively capture the dynamic spatial-temporal dependencies. Besides, we adopt a multi-subgraph encoding mechanism to enhance the representation of complex spatial dependency. Extensive experiments on two real-world datasets demonstrate that the performance of our MAGCN exceeds the state-of-the-art baselines.

Using fixed rate uniform vector quantization, in this paper, we consider how to design a joint watermarking and compression (JWC) system for Gaussian and Laplacian sources to maximize the robustness in the presence of additive Gaussian attacks under constraints on the compression rate and quantization distortion. Firstly, we construct vector quantizers shaped to match the multidimensional distribution of source signals. Then we scale codebooks corresponding to the vector quantizers to maximize the robustness of the watermarks against the additive Gaussian attacks. Simulation results show that the proposed scheme can achieve up to 0.92 dB distortion-to-noise ratio (DNR) gain over JWC schemes using uniform scalar quantization while maintaining the simplicity of implementation with uniform quantization.

We propose the concept of quality-aware image, in which certain extracted features of the original (high-quality) image are embedded into the image data as invisible hidden messages. When a distorted version of such an image is received, users can decode the hidden messages and use them to provide an objective measure of the quality of the distorted image. To demonstrate the idea, we build a practical quality-aware image encoding, decoding and quality analysis system, which employs: 1) a novel reduced-reference image quality assessment algorithm based on a statistical model of natural images and 2) a previously developed quantization watermarking-based data hiding technique in the wavelet transform domain.

In this paper we present optimal quantization watermarking strategies with respect to the robustness of a watermarking system given the embedding rate and distortion constraint. Firstly, we investigate the optimal decoding for quantization watermarking and show that by making use of channel statistics, the maximum likelihood decoder is always better than the minimum distance decoder. Secondly, the optimal encoding is designed by exploiting the knowledge of the host signal and channel statistics. Algorithms for designing the optimal uniform quantization encoding scheme and optimal nonuniform quantization encoding scheme are proposed. Simulation results show that the optimal nonuniform quantization watermarking can achieve better performance. Finally, applications to image authentication which is robust to high quality JPEG compression are described.