Road extraction from remote sensing images is of great significance to urban planning, navigation, disaster assessment, and other applications. Although deep neural networks have shown a strong ability in road extraction, it remains a challenging task due to complex circumstances and factors such as occlusion. To improve the accuracy and connectivity of road extraction, we propose an inner convolution integrated encoder-decoder network with the post-processing of directional conditional random fields. Firstly, we design an inner convolutional network which can propagate information slice-by-slice within feature maps, thus enhancing the learning of road topology and linear features. Additionally, we present the directional conditional random fields to improve the quality of the extracted road by adding the direction of roads to the energy function of the conditional random fields. The experimental results on the Massachusetts road dataset show that the proposed approach achieves high-quality segmentation results, with the F1-score of 84.6%, which outperforms other comparable “state-of-the-art” approaches. The visualization results prove that the proposed approach is able to effectively extract roads from remote sensing images and can solve the road connectivity problem produced by occlusions to some extent.

In recent years, object detection using aerial images captured by unmanned aerial vehicles (UAVs) has become a research hotspot. However, due to the high resolution of UAV imagery, large variations in object scales, and the predominance of small targets, achieving fast and accurate object classification remains a significant challenge. To address these issues, this paper proposes a lightweight feature enhancement attention network, termed LFEA-YOLO. First, a multi-scale feature fusion (MSFF) module is introduced, which establishes a cross-scale feature interaction mechanism to effectively integrate semantic information with spatial details while reducing the computational overhead caused by redundant features. Second, deformable convolution networks (DCNs) are incorporated into the neck network to construct the DCNv4-C2F module. Leveraging its adaptive spatial sampling capability, this module dynamically adjusts the receptive field of convolution kernels, thereby overcoming the limitations of traditional fixed-grid sampling and enhancing the network’s ability to model geometric variations of small-scale objects. Finally, an attention-enhanced detection head (AE-Head) is designed, integrating a large separable kernel attention (LSKA) mechanism to dynamically emphasize discriminative features in key target regions. Experimental results on the VisDrone2019 and DOTAv1.0 datasets demonstrate significant improvements over the baseline model, with ${\mathrm{mAP}}_{50}$ and ${\mathrm{mAP}}_{50:95}$ increasing by 10.6% and 7.8%, respectively. Compared with several recently proposed models, LFEA-YOLO also exhibits clear advantages, validating the effectiveness of the proposed approach.

RGBT (visible-thermal) object tracking holds significant value in complex scenarios such as low-light and hazy environments, enabling robust all-weather tracking by leveraging the complementary strengths of visible and thermal infrared modalities. However, challenges such as target appearance variations, similar object interference, and camera motion often lead to tracking drift. This paper proposes RecheckTrack, a robust RGBT tracking framework that addresses these issues through the enhancement of temporal information and a backward trajectory verification mechanism. The dual-branch fusion network adaptively learns target dynamics using appearance tokens and modality tokens. Modality tokens focus on high-quality features and target-probable regions, while appearance tokens track dynamic changes in target appearance, improving robustness against deformation, occlusion, and scale variations. To mitigate drift caused by sudden target or camera motion, a recheck network is introduced, which employs a two-stage candidate box selection method and jointly matches targets using bidirectional tracking consistency and appearance similarity. Additionally, for long-term tracking scenarios where targets may be lost, the recheck network is improved with a path-consistency-based backward trajectory selection method and an approximate global search strategy, efficiently recovering lost targets. Experiments on the VTUAV, LasHeR, and RGBT234 datasets demonstrate that RecheckTrack significantly reduces tracking drift and improves accuracy, providing an effective solution for RGBT tracking in complex scenarios.

Large Language Models (LLMs) have demonstrated remarkable performance across diverse domains, thereby prompting researchers to explore their potential for use in recommendation systems. Initial at-tempts have leveraged the exceptional capabilities of LLMs, such as rich knowledge and strong gener-alization through In-context Learning, which involves phrasing the recommendation task as prompts.Nevertheless, Given the significant disparity between the training tasks of LLMs and recom-mendation tasks, along with the data sparsity issue arising from insufficient recommendation data during pre-training, the performance of LLMs in recommendation tasks remains unsatisfactory. To bridge this gap, we propose fine-tuning LLMs with Contrastive Learning-optimized recommendation data to Construct a Language Model Recommendation. namely CLCLMRec. The CLCLMRec frame-work aims to achieve an effective and efficient integration of Large Language Models with recommen-dation systems in settings with low GPU memory consumption. The framework encompasses two cru-cial tuning phases: firstly, instruction tuning, a universal LLM training process designed to enhance the generalization capabilities of LLMs; and secondly, recommendation tuning, which focuses on fine-tuning the model specifically for recommendation tasks. In Rec-tuning, we employ the self-attention mechanism to process embedded data and emphasize sequential recommendation methods to capture users' dynamic interests. However, addressing the challenge of data sparsity, we innovatively intro-duce contrastive learning, extracting self-supervised signals from user behavior sequences to train large models more effectively and uncover meaningful user patterns. Ultimately, we fuse "recommen-dation instructions" with "recommendation inputs" to construct "instruction inputs" and utilize "rec-ommendation outputs" as "instruction outputs" for training, thereby building a Large Recommendation Language Model (LRLM). Experimental results demonstrate that even with extremely limited data (fewer than 100 samples), the CLCLMRec framework can significantly boost the recommendation ca-pabilities of LLMs in domains such as movies and books.

Deep joint source-channel coding (DeepJSCC) has emerged as a promising paradigm for semantic-aware wireless image transmission, achieving strong performance under challenging channel conditions. However, MSE-trained DeepJSCC systems typically achieve high peak signal-to-noise ratio (PSNR) values but suppress high-frequency details, resulting in perceptually blurry reconstructions that fail to capture fine textures and edge information. Existing perceptual enhancement approaches for JSCC systems face significant practical limitations: full transceiver redesign methods require replacing both the transmitter and the receiver with large models (19–31 million parameters), incurring substantial deployment costs; diffusion-based refinement approaches require over 1700 million additional parameters and introduce inference latency exceeding 13 s, rendering them unsuitable for latency-constrained wireless applications; and generic image restoration networks lack channel state awareness and cannot adapt to varying signal-to-noise ratio (SNR) conditions. This paper proposes a lightweight receiver-only perceptual enhancer designed for use with frozen DeepJSCC backbones. The proposed module adopts residual learning with feature-wise linear modulation (FiLM)-based SNR-adaptive modulation to dynamically adjust the enhancement strength under varying channel conditions. A radially weighted FFT magnitude loss is further introduced to guide high-frequency recovery. The enhancer adds only 0.29 million trainable parameters (<1% of the backbone) and requires neither transmitter modification nor backbone retraining. Extensive experiments on the Kodak24 and DIV2K datasets demonstrate a 34.4–37.5% LPIPS reduction over the frozen DeepJSCC baseline under AWGN channels. Supplementary robustness evaluations further show a 30–33% LPIPS reduction under Rayleigh fading, and stable generalization to unseen SNR levels. The receiver-side decoder-plus-enhancer pipeline requires 43 ms at 768 × 512 resolution, corresponding to approximately 23 frames per second.