Ziqi Chen

Video stabilization is a critical technology for enhancing visual content quality in dynamic shooting scenarios, especially with the widespread adoption of mobile photography devices and Unmanned Aerial Vehicle (UAV) platforms. While traditional digital stabilization algorithms can improve frame stability by modeling global motion trajectories, they often suffer from excessive cropping or boundary distortion, leading to a significant loss of valid image regions. To address this persistent challenge, we propose the View Out-boundary Synthesis Algorithm (VOSA), a symmetry-aware spatio-temporal consistency framework. By leveraging rotational and translational symmetry principles in motion dynamics, VOSA realizes optical flow field extrapolation through an encoder–decoder architecture and an iterative boundary extension strategy. Experimental results demonstrate that VOSA enhances conventional stabilization by increasing content retention by 6.3% while maintaining a 0.943 distortion score, outperforming mainstream methods in dynamic environments. The symmetry-informed design resolves stability–content conflicts and outperforms mainstream methods in dynamic environments, establishing a new paradigm for full-frame stabilization.

Distantly Supervised Relation Extraction (DSRE) aims to automatically identify semantic relationships within large text corpora by aligning with external knowledge bases. Despite the success of current methods in automating data annotation, they introduce two main challenges: label noise and data long-tail distribution. Label noise results in inaccurate annotations, which can undermine the quality of relation extraction. The long-tail problem, on the other hand, leads to an imbalanced model that struggles to extract less frequent, long-tail relations. In this paper, we introduce a novel relation extraction framework based on multi-level hierarchical attention. This approach utilizes Graph Attention Networks (GATs) to model the hierarchical structure of the relations, capturing the semantic dependencies between relation types and generating relation embeddings that reflect the overall hierarchical framework. To improve the classification process, we incorporate a multi-level classification structure guided by hierarchical attention, which enhances the accuracy of both head and tail relation extraction. A local probability constraint is introduced to ensure coherence across the classification levels, fostering knowledge transfer from frequent to less frequent relations. Experimental evaluations on the New York Times (NYT) dataset demonstrate that our method outperforms existing baselines, particularly in the context of long-tail relation extraction, offering a comprehensive solution to the challenges of DSRE.

Polyp segmentation is crucial for early colorectal cancer detection, but accurately delineating polyps is challenging due to their variations in size, shape, and texture and low contrast with surrounding tissues. Existing methods often rely solely on spatial-domain processing, which struggles to separate high-frequency features (edges, textures) from low-frequency ones (global structures), leading to suboptimal segmentation performance. We propose the Dynamic Frequency-Decoupled Refinement Network (DFDRNet), a novel segmentation framework that integrates frequency-domain and spatial-domain processing. DFDRNet introduces the Frequency Adaptive Decoupling (FAD) module, which dynamically separates high- and low-frequency components, and the Frequency Adaptive Refinement (FAR) module, which refines these components before fusing them with spatial features to enhance segmentation accuracy. Embedded within a U-shaped encoder–decoder framework, DFDRNet achieves state-of-the-art performance across three benchmark datasets, demonstrating superior robustness and efficiency. Our extensive evaluations and ablation studies confirm the effectiveness of DFDRNet in balancing segmentation accuracy with computational efficiency.

Semantic segmentation of high-resolution remote sensing images (HRRSIs) presents unique challenges due to the intricate spatial and spectral characteristics of these images. Traditional methods often prioritize spatial information while underutilizing the rich spectral context, leading to limited feature discrimination capabilities. To address these issues, we propose a novel frequency attention-enhanced network (FAENet), which incorporates a frequency attention model (FreqA) to jointly model spectral and spatial contexts. FreqA leverages discrete wavelet transformation (DWT) to decompose input images into distinct frequency components, followed by a two-stage attention mechanism comprising inner-component channel attention (ICCA) and cross-component channel attention (CCCA). These mechanisms enhance spectral representation, which is further refined through a self-attention (SA) module to capture long-range dependencies before transforming back into the spatial domain. FAENet’s encoder–decoder architecture facilitates multiscale feature refinement, enabling effective segmentation. Extensive experiments on the ISPRS Potsdam and LoveDA benchmarks demonstrate that FAENet outperforms state-of-the-art models, achieving superior segmentation accuracy. Ablation studies further validate the contributions of ICCA and CCCA. Moreover, efficiency comparisons confirm the superiority of the proposed FAENet over other models.

Colorectal cancer remains a leading cause of cancer-related deaths worldwide, with early detection and removal of polyps being critical in preventing disease progression. Automated polyp segmentation, particularly in colonoscopy images, is a challenging task due to the variability in polyp appearance and the low contrast between polyps and surrounding tissues. In this work, we propose an edge-enhanced network (EENet) designed to address these challenges by integrating two novel modules: the covariance edge-enhanced attention (CEEA) and cross-scale edge enhancement (CSEE) modules. The CEEA module leverages covariance-based attention to enhance boundary detection, while the CSEE module bridges multi-scale features to preserve fine-grained edge details. To further improve the accuracy of polyp segmentation, we introduce a hybrid loss function that combines cross-entropy loss with edge-aware loss. Extensive experiments show that the EENet achieves a Dice score of 0.9208 and an IoU of 0.8664 on the Kvasir-SEG dataset, surpassing state-of-the-art models such as Polyp-PVT and PraNet. Furthermore, it records a Dice score of 0.9316 and an IoU of 0.8817 on the CVC-ClinicDB dataset, demonstrating its strong potential for clinical application in polyp segmentation. Ablation studies further validate the contribution of the CEEA and CSEE modules.

Semantic segmentation plays a pivotal role in the intelligent interpretation of remote sensing images (RSIs). However, conventional methods predominantly focus on learning representations within the spatial domain, often resulting in suboptimal discriminative capabilities. Given the intrinsic spectral characteristics of RSIs, it becomes imperative to enhance the discriminative potential of these representations by integrating spectral context alongside spatial information. In this paper, we introduce the spectrum-space collaborative network (SSCNet), which is designed to capture both spectral and spatial dependencies, thereby elevating the quality of semantic segmentation in RSIs. Our innovative approach features a joint spectral–spatial attention module (JSSA) that concurrently employs spectral attention (SpeA) and spatial attention (SpaA). Instead of feature-level aggregation, we propose the fusion of attention maps to gather spectral and spatial contexts from their respective branches. Within SpeA, we calculate the position-wise spectral similarity using the complex spectral Euclidean distance (CSED) of the real and imaginary components of projected feature maps in the frequency domain. To comprehensively calculate both spectral and spatial losses, we introduce edge loss, Dice loss, and cross-entropy loss, subsequently merging them with appropriate weighting. Extensive experiments on the ISPRS Potsdam and LoveDA datasets underscore SSCNet’s superior performance compared with several state-of-the-art methods. Furthermore, an ablation study confirms the efficacy of SpeA.

Contextual information plays a pivotal role in the semantic segmentation of remote sensing imagery (RSI) due to the imbalanced distributions and ubiquitous intra-class variants. The emergence of the transformer intrigues the revolution of vision tasks with its impressive scalability in establishing long-range dependencies. However, the local patterns, such as inherent structures and spatial details, are broken with the tokenization of the transformer. Therefore, the ICTNet is devised to confront the deficiencies mentioned above. Principally, ICTNet inherits the encoder–decoder architecture. First of all, Swin Transformer blocks (STBs) and convolution blocks (CBs) are deployed and interlaced, accompanied by encoded feature aggregation modules (EFAs) in the encoder stage. This design allows the network to learn the local patterns and distant dependencies and their interactions simultaneously. Moreover, multiple DUpsamplings (DUPs) followed by decoded feature aggregation modules (DFAs) form the decoder of ICTNet. Specifically, the transformation and upsampling loss are shrunken while recovering features. Together with the devised encoder and decoder, the well-rounded context is captured and contributes to the inference most. Extensive experiments are conducted on the ISPRS Vaihingen, Potsdam and DeepGlobe benchmarks. Quantitative and qualitative evaluations exhibit the competitive performance of ICTNet compared to mainstream and state-of-the-art methods. Additionally, the ablation study of DFA and DUP is implemented to validate the effects.

Semantic segmentation has been a fundamental task in interpreting remote sensing imagery (RSI) for various downstream applications. Due to the high intra-class variants and inter-class similarities, inflexibly transferring natural image-specific networks to RSI is inadvisable. To enhance the distinguishability of learnt representations, attention modules were developed and applied to RSI, resulting in satisfactory improvements. However, these designs capture contextual information by equally handling all the pixels regardless of whether they around edges. Therefore, blurry boundaries are generated, rising high uncertainties in classifying vast adjacent pixels. Hereby, we propose an edge distribution attention module (EDA) to highlight the edge distributions of leant feature maps in a self-attentive fashion. In this module, we first formulate and model column-wise and row-wise edge attention maps based on covariance matrix analysis. Furthermore, a hybrid attention module (HAM) that emphasizes the edge distributions and position-wise dependencies is devised combing with non-local block. Consequently, a conceptually end-to-end neural network, termed as EDENet, is proposed to integrate HAM hierarchically for the detailed strengthening of multi-level representations. EDENet implicitly learns representative and discriminative features, providing available and reasonable cues for dense prediction. The experimental results evaluated on ISPRS Vaihingen, Potsdam and DeepGlobe datasets show the efficacy and superiority to the state-of-the-art methods on overall accuracy (OA) and mean intersection over union (mIoU). In addition, the ablation study further validates the effects of EDA.

Pan-sharpening is a significant task that aims to generate high spectral- and spatial- resolution remote-sensing image by fusing multi-spectral (MS) and panchromatic (PAN) image. The conventional approaches are insufficient to protect the fidelity both in spectral and spatial domains. Inspired by the robust capability and outstanding performance of convolutional neural networks (CNN) in natural image super-resolution tasks, CNN-based pan-sharpening methods are worthy of further exploration. In this paper, a novel pan-sharpening method is proposed by introducing a multi-scale channel attention residual network (MSCARN), which can represent features accurately and reconstruct a pan-sharpened image comprehensively. In MSCARN, the multi-scale feature extraction blocks comprehensively extract the coarse structures and high-frequency details. Moreover, the multi-residual architecture guarantees the consistency of feature learning procedure and accelerates convergence. Specifically, we introduce a channel attention mechanism to recalibrate the channel-wise features by considering interdependencies among channels adaptively. The extensive experiments are implemented on two real-datasets from GaoFen series satellites. And the results show that the proposed method performs better than the existing methods both in full-reference and no-reference metrics, meanwhile, the visual inspection displays in accordance with the quantitative metrics. Besides, in comparison with pan-sharpening by convolutional neural networks (PNN), the proposed method achieves faster convergence rate and lower loss.