Xingquan Cai

Marine oil spills pose serious threats to coastal ecosystems and maritime activities, and synthetic aperture radar (SAR) has become an important tool for all-weather marine monitoring. However, SAR oil spill detection remains challenging because oil spills usually appear as weak dark anomalies with blurred boundaries, elongated or fragmented shapes, and strong interference from lookalike phenomena such as low-wind areas and internal waves. To address these issues, we propose RGCNet, a lightweight radiometric–geometric–contextual detection framework based on YOLOv11n. Firstly, the H_SPDRFF module is incorporated into the backbone to enhance weak radiometric responses through constrained feature amplification, thereby reducing missed detections caused by low-contrast oil slicks. Secondly, the C3k2_GSR module is designed in the neck to strengthen anisotropic geometric refinement and preserve the continuity of elongated and fragmented oil spill regions during multi-scale feature fusion. Finally, a SAR-adapted large selective kernel block (LSKBlock) is embedded in the high-level backbone to improve contextual discrimination between true oil spills and lookalike dark formations. Experiments on DeepSAR show that RGCNet increases mAP@0.5 and mAP@0.5:0.95 by 3.6 and 3.0 percentage points over the YOLOv11n baseline, respectively. Cross-dataset evaluation on SAR-Oil-Spill demonstrates a 3.9-point mAP@0.5 gain, indicating strong transferability. Furthermore, with a compact model size of 2.67 M parameters and 6.4 G FLOPs, RGCNet achieves an inference speed of 162.5 FPS on an RTX A4000 GPU, demonstrating its efficiency and potential for real-time maritime surveillance. Nevertheless, the current bounding-box formulation cannot precisely delineate irregular oil-spill boundaries. Future work will therefore investigate fine-grained segmentation and cross-sensor adaptation.

Multimodal emotion recognition in conversation (MERC) plays a crucial role in applications such as empathetic dialogue systems, intelligent tutoring, and mental health monitoring. Existing graph neural network (GNN)-based approaches often rely on fixed conversational topologies, which may introduce redundant message passing and fail to capture evolving high-order dependencies. In addition, conventional GNNs model temporal dynamics in discrete steps, leading to oversmoothing and a limited ability to represent continuous emotional changes. To address these limitations, we propose CAS-ODE (Contrastive Adaptive Spatiotemporal ODE Network), a novel framework that jointly learns adaptive conversational structures and continuous temporal dynamics. CAS-ODE employs a variational hypergraph autoencoder with contrastive regularization to infer robust high-order relations, while a graph ordinary differential equation models the smooth evolution of emotional states. Extensive experiments on the IEMOCAP and MELD benchmarks validate the superiority of CAS-ODE, which achieves state-of-the-art performance and highlights the effectiveness of synergistically learning conversational structures and their continuous temporal evolution.

Pedestrian and vehicle in adverse weather detection technology based on deep learning has attracted increasing attention. However, YOLO-based detectors degrade under low contrast and noise, causing missed small or occluded targets, while DETR-based models suffer from slow convergence and limited multi-scale sensitivity. To address these challenges, a robust weather-aware transformer for vehicle and pedestrian detection, RWT-DETR, is proposed. Firstly, a Strip-Oriented Feature Gating (SOFG) module is designed to improve the capture of strip-like and directional features of traffic targets. Secondly, a Signed Attention Decomposition (SAD) mechanism is designed to enable the encoder to model both positive and negative correlations, thereby enhancing the model's robustness against noise interference and feature blurring. Finally, an Adaptive Multi-scale Fusion with Gating (AMFG) module is proposed to achieve dynamic calibration and denoising of cross-scale features, significantly improving semantic alignment. Experimental results show that, compared with the baseline, the proposed method improves mAP@0.5 and mAP@0.5:0.95 by 2.4% and 3.0%, respectively. Meanwhile, it outperforms existing mainstream models, confirming its effectiveness under adverse weather conditions.

Encrypted traffic classification is a fundamental task in cybersecurity and network traffic management. With the popularity of deep learning, many studies have used it to achieve high accuracy in this field. However, these models rely on large amounts of labeled data and require retraining or fine-tuning when faced with new tasks, which is time-consuming and resource-intensive. Due to this limitation, few-shot traffic classification methods have emerged to reduce dependence on large labeled datasets by using only a few samples. However, these models often still require pre-training and fine-tuning to perform well. Meanwhile, Large Language Models (LLMs), such as ChatGPT, have attracted attention for their capacity to learn new concepts during inference without fine-tuning. Inspired by this, we propose In-Context Encrypted Traffic, a novel few-shot network traffic classification method allowing for few-shot classification without the need for fine-tuning. We use two existing encrypted traffic datasets to demonstrate the effectiveness of our method when facing new samples. We evaluate our method on a realistic 5-way few-shot classification task over encrypted traffic, and it achieves up to 5.7% higher accuracy than representative meta-learning baselines, without requiring fine-tuning.

Deep learning-based Synthetic Aperture Radar (SAR) despeckling faces two key challenges: physical model mismatch, where simplified noise assumptions fail to capture speckle’s statistics, and architectural generality, where standard networks struggle to separate noise from texture. To address this, we propose DA-PhysDiff, a dual-aware physics-informed diffusion model for SAR image despeckling. The proposed LogGamma-based stochastic differential equation models multiplicative speckle more accurately, while the dual-aware network composed of a Statistics-Conditioned Gamma Attention (SCGA) and a Frequency-Refined Adaptive Wavelet Convolution (FRA-WTConv) enhances both distribution and frequency awareness. Experiments on HRSID, MiniSAR, and synthetic benchmarks demonstrate that DA-PhysDiff achieves up to +2.5 dB PSNR and +4.8% EPD-ROA improvement over recent diffusion models, while reducing parameters by 19% compared with SAR-DDPM and maintaining real-time inference.