In modern robotic vision-based grasping tasks, accurate object detection, localization, and grasp execution are crucial. To improve grasping success rates and efficiency, we propose a vision-based robotic grasping system designed to address rapid object recognition and grasping in cluttered scenarios. The system integrates an Attention-Ghost(AG) mechanism, a Multi-Parallel Convolution (MPC) module, a Depthwise Separable Convolution (DSC) module, and residual connections. To validate its effectiveness, extensive experiments were conducted. Results show that our method achieves accuracies of 98.8

Diffusion-based policies have shown strong performance in robot imitation learning, but directly adopting diffusion processes from image generation introduces mismatches with the low-dimensional, heterogeneous, and latency-sensitive nature of robot action spaces. To address this, we revisit the entire generative pipeline and propose the Full-process Adapted Diffusion Policy (FADP). Specifically, in the forward process, we analyze the widely used cosine noise schedule through a mode-separability framework, revealing its effectiveness in preserving mode discriminability in low-dimensional regimes and its interaction with subsequent processes. In the reverse process, we replace the shared inverse variance in standard DDPM with a learnable per-dimension formulation to better capture heterogeneous joint dynamics. In the sampling process, we employ consistency distillation to compress multi-step denoising into few-step inference, reducing latency while maintaining stability. These improvements require no architectural modifications and incur negligible overhead. Extensive experiments on both simulated and real-world robotic tasks demonstrate that FADP outperforms existing baselines and achieves competitive performance, highlighting the importance of full-process adaptation for diffusion-based robot policies.

Accurate detection and counting of apple flowers are essential for digital orchard monitoring and intelligent flower-fruit management. However, apple flowers in natural orchards are typically small, densely distributed, and subject to frequent occlusion, making conventional detection approaches unreliable. To address these challenges, we develop a Density-guided Dual-stage High-resolution Detection framework (DG-DSHRD) combined with an enhanced YOLO-FRNet architecture. The proposed DG-DSHRD performs coarse detection to produce a density map and adaptively determines high-density regions for refined inference, reducing redundant computation while preserving structural continuity. YOLO-FRNet further improves small-object detection through an enhanced backbone, a lightweight multi-scale feature pyramid, and a quality-aware detection head, enabling more robust representation and localization under complex orchard backgrounds. Experiments conducted on a self-built distant-view apple flower detection dataset demonstrate that the combined framework effectively mitigates background interference, enhances feature fusion, and outperforms mainstream detectors and the slicing aided hyper inference(SAHI)-based static slicing. Compared with YOLOv11s using SAHI-based static slicing, the proposed YOLOv11s-FRNet + DG-DSHRD framework improves mAP50 and F1 by 6.2 and 3.1 percentage points, respectively, reaching 66.6 % and 66.3 %, while increasing the inference speed from 2.85 FPS to 10.3 FPS, thereby enabling near real-time deployment. Overall, the DG-DSHRD + YOLO-FRNet framework provides an accurate, efficient, and deployable solution for high-density small-object detection in orchard environments, offering strong support for automated flower counting and intelligent flower-fruit regulation.

Small object detection is a critical and challenging task in UAV applications due to limited pixel information and feature degradation in deep neural networks. To address these issues, this paper proposes a novel object detector, the Multi-Scale and Multi-Receptive Field Network (MS-MRFNet). A Multi-Scale Weighted Feature Pyramid Network (MSWFPN) is designed to enhance feature fusion across scales by adaptively weighting semantic and spatial features. Additionally, a Multi-Channel Feature Aggregation (MCFA) module is developed to aggregate features at the same level, employing reparameterization trick to retain critical information. Finally, Wise-IoU (WIoU) is adopted as the bounding box regression loss, balancing gradient gains from samples of varying quality across scales and feature groups, thereby accelerating model convergence. Experimental results on the VisDrone 2019, TinyPerson, and NWPU VHR-10 datasets demonstrate mAP50 improvements of 8.6

Zero-shot hand gesture recognition aims to recognize unseen gesture categories without labeled training samples by transferring knowledge from seen classes. However, common class representations, such as category names or generic textual descriptions, are often too coarse to capture the fine-grained geometric characteristics of hand gestures, weakening visual-semantic alignment and unseen-class generalization. To address this issue, we propose a geometry-aware semantic modeling framework that constructs class prototypes from explicit geometry-aware descriptions that encode finger extension patterns, relative finger configurations, and overall hand shape properties. A lightweight shared prompt adaptation module is introduced to refine these prototypes and improve visual-semantic alignment in a parameter-efficient manner. Experiments on HaGRID under both ZSL and GZSL settings show consistent gains over class-name and generic-description baselines. Qualitative results also reveal fewer confusions among visually similar unseen gestures, supporting the use of geometry-driven class descriptions for gesture zero-shot recognition.

Diabetic Retinopathy (DR) and Hypertensive Retinopathy (HR) are major retinal diseases contributing to global visual impairment, where early screening and progression monitoring critically depend on accurate retinal Artery/Vein segmentation and diameter estimation. To address limitations in current methods regarding vessel segmentation and geometric quantification, we propose a framework integrating both segmentation and measurement. The architecture includes three novel deep learning modules: the Hierarchical Feature Extraction and Integration Module (HFEIM) for capturing multi-scale vessel structures, the Adaptive Attention Pooling Module (AAPM) for emphasizing critical vascular regions, and the Multi-Scale Attention Residual Enhancement Module (MSAREM) to enhance detection of fine vessels. Following precise arteriovenous segmentation, we apply skeletonization to extract vessel centerlines, and reconstruct orthogonal elliptical cross-sections along the vessels. The final diameter is derived using the area-equivalent principle, ensuring physiological accuracy. Our approach demonstrates excellent segmentation accuracy and robustness across various arteriovenous segmentation datasets, offering a powerful solution to current challenges in retinal vessel segmentation and retinal diseases detection, these results demonstrate the effectiveness of the proposed model.

Computer-Aided Tongue Diagnosis (CATD) is a core component of intelligent Traditional Chinese Medicine (TCM). However, the extreme scarcity of high-quality datasets with fine-grained pathological labels (e.g., tooth marks, cracks) severely limits the generalization capability and clinical diagnostic accuracy of downstream segmentation models. Our previous tongue data augmentation method (the GAN-based T-STGN) is prone to mode collapse when processing complex edges and high-frequency textures, leading to insufficient realism in the generated images. In this paper, we propose a novel frequency-aware two-stage mask-guided diffusion model, FaMD, which decouples the generation process into two stages: “morphology construction” and “texture rendering”. In the first stage, anatomical morphology masks are generated based on clinical pathological labels to provide strict boundary priors for tongue image synthesis. In the second stage, three major innovations are introduced during the image denoising diffusion process: 1) To address the high-frequency detail loss caused by spatial compression, a Time-Gated Residual Wavelet Downsampling (TG-ResLWD) module is designed to dynamically protect fine pathological features; 2) A Frequency-Aware Spatially-Adaptive Normalization (Fa-SPADE) is proposed to achieve decoupled modulation and precise alignment between macroscopic anatomical morphology masks and microscopic pathological details; 3) A Dynamic Fast Fourier Transform (FFT) Loss is introduced to impose global structural constraints in the complex frequency domain, effectively mitigating the frequency aliasing problem. Experiments on a 1.1k real tongue image dataset demonstrate that the generation quality of FaMD significantly outperforms existing state-of-the-art (SOTA) methods such as T-STGN and LDM, reducing the FID to 53.74. When the FaMD synthesized data is utilized for training augmentation, the Dice indices of the classic U-Net and DeepLabv3+ segmentation models improve to 99.09

High-speed UAV images challenge real-time small-object detection because directional motion blur weakens already fragile object evidence, disrupts cross-scale feature transport, and aggravates interference from cluttered backgrounds. To address this issue, we propose MAF-Det, a D-Fine-S-based detector that strengthens the backbone–neck feature path with three lightweight modules: Blur-Aligned Spatial Aggregation (BASA) for deformable large-kernel aggregation on projected multi-scale features, Adaptive Alignment Resampling (AAR) for offset-guided top-down resampling, and Selective Conflict Modulation (SCM) for channel-adaptive fusion of propagated and lateral features. Instead of relying on image restoration or a heavier detection head, MAF-Det focuses on stabilizing aggregation, resampling, and fusion under blur. On VisDrone2019, under both clear images and direction-dependent synthetic motion blur, MAF-Det improves from 0.101 to 0.158 on the blurred test set while maintaining 37.1 FPS on Jetson Orin Nano 4G with TensorRT FP16.

Zero-shot learning (ZSL) enables recognition of classes that have no labeled training examples by exploiting semantic descriptions such as attributes. Among different ZSL paradigms, embedding-based methods perform recognition by projecting image representations toward semantic prototypes. However, their performance can be limited when a simple projection head cannot sufficiently adapt visual features to the semantic prototype space. Motivated by this observation, we introduce ViT-CARP, which couples a vision-language pretrained CLIP-ViT encoder with the proposed Channel-Attentive Residual Projection (CARP) module within an embedding-based ZSL framework. CARP formulates semantic projection as a channel-aware feature refinement process. It explicitly recalibrates feature channels to suppress task-irrelevant noise while preserving discriminative visual information through residual learning. Together, these designs produce attribute-aware embeddings for prototype-based recognition. Experiments on AWA2, SUN, and CUB show that ViT-CARP achieves strong performance among recent embedding-based baselines under both conventional ZSL and generalized ZSL settings. These results verify that channel-level projection refinement improves visual-semantic alignment for zero-shot recognition.

This paper proposes a method for controlling IAP (Integrated Aerial Platform) drones based on LLMs (large language models), including both API (Application Programming Interface) calls and locally deployed models, with the aim of generating real-time control commands through natural language input. The system adopts a modular design, consisting of three modules: the Natural Language Instruction Parsing Module, the Control Command Generation Module, and the Admittance Control Module. These modules work in close collaboration to achieve natural language-based human-drone interaction (HDI). To address issues related to cloud-based LLMs, such as network fluctuations and the reliability of generated code, we propose a solution combining both cloud-based and local deployment of LLMs, and validate the generated code through a safety check mechanism. Finally, a series of experiments in the RaiSim simulation environment validate the system’s effectiveness in performing both basic flight tasks and complex peg-in-hole tasks. The experimental results show that the system can effectively generate control code and, by integrating drone pose, external force, and torque feedback, collaborate with the control module to complete high-precision tasks, demonstrating the great potential of using LLMs for natural language interaction with drones.

While existing knowledge distillation (KD) methods typically force students to mimic teacher features without considering prediction reliability, this practice risks propagating the teacher’s erroneous supervision to the student. To address this, we propose the Logitwise Distillation Network (LDN), a novel framework that dynamically quantifies sample-wise confidence through the ranking of ground truth labels in teacher logits. Specifically, LDN introduces three key innovations: (1) weighted class means that prioritize high-confidence samples, (2) adaptive feature selection based on logit ranking, and (3) positive–negative sample adjustment (PNSA) to reverse error-prone supervision. These components are unified into a feature direction (FD) loss, which guides students to selectively emulate trustworthy teacher features. Experiments on CIFAR-100 and ImageNet demonstrate that LDN achieves state-of-the-art performance, improving accuracy by 0.3–0.5% over SOTA methods. Notably, LDN exhibits stronger compatibility with homogeneous networks (2.4% gain over baselines) and requires no additional training costs when integrated into existing KD pipelines. This work advances feature distillation by addressing error propagation, offering a plug-and-play solution for reliable knowledge transfer.

Tiny object detection (TOD) reveals a fundamental flaw in feature pyramid networks: high-level features (P5-P6) frequently receive zero positive anchors under standard label assignment protocols, leaving their semantic representations untrained due to exclusion from loss computation. This creates dual deficiencies: (1) Stranded high-level features become semantic dead-ends without gradient updates, while (2) low-level features lack essential semantic context for robust classification. We propose E-FPN-BS that systematically converts wasted high-level semantics into low-level feature enhancements. To address these issues, we propose E-FPN-BS, a novel architecture integrating multi-scale feature enhancement and adaptive optimization. First, our Context Enhancement Module(CEM) employs dual-branch processing to align and compress high-level features for effective global-local fusion. Second, the Foreground-Background Separation Module (FBSM) generates spatial gating masks that dynamically amplify discriminative regions. To address gradient imbalance across object scales, we further propose a Dynamic Gradient-Balanced Loss (DCLoss) that automatically modulates loss contributions via scale-aware gradient equilibrium. Extensive experiments across multiple benchmark datasets demonstrate the outstanding performance and generalization ability of our approach.

Current regression losses for tiny object detection exhibit fundamental limitations in error sensitivity control. While L1 loss maintains constant gradients regardless of error magnitude (impeding high-precision convergence), L2 loss suffers from outlier sensitivity due to its quadratic growth. Though Smooth L1 introduces threshold-based L1/L2 switching, its fixed transition boundary fails to address the critical need for adaptive gradient modulation, particularly vital for small objects where minor deviations significantly impact IoU calculations. To address the limitations of existing regression losses, we propose a Gradient-Adaptive Hyperbolic Loss (GAHL) that intrinsically couples gradient behaviors with error scales. Specifically, GAHL loss first overcomes the weak supervision of L1 near convergence through hyperbolic gradient amplification, which progressively intensifies updates for diminishing errors, achieving several times stronger gradients than L1 for sub-pixel deviations critical to IoU sensitivity. Then by replacing the rigid threshold in Smooth L1 with error-adaptive blending of L1/L2 regimes, it eliminates oscillatory optimization, reducing convergence instability on sub-16-pixel targets. Simultaneously, controlled saturation mechanics suppress large geometrically irrelevant errors, capping their gradient magnitudes at 70

Traditional gesture-based human-robot interaction relies on one-to-one gesture-command mapping, requiring numerous gestures and imposing high cognitive load. Existing networks are often computationally heavy, limiting real-time deployment on resource-constrained robots. To significantly reduce the required gestures and improve the intuitiveness of user experience, we develop the Spatial Semantic Mapping framework to change the gesture-based control paradigm by assigning commands based on the spatial position of the hand, establishing a flexible one-to-many mapping. To achieve an optimal balance between accuracy and computational efficiency, we propose a Lightweight Multi-level Fusion Segmentation Network (LMFSNet). Firstly, to reduce computational costs greatly, we propose a lightweight Residual Axial Group Convolution as the core operation of the model. Secondly, to maintain high performance in the lightweight network, we design two modules: Dynamic Adaptive Attention Block (DAAB) and Long-Short Distance Extraction (LSDE) block. Specifically, the DAAB dynamically reweights features to focus on important information, and the LSDE effectively captures and fuses multi-scale features. Experimental results show that the proposed LMFSNet achieves state-of-the-art accuracy while maintaining real-time speed and a compact model size.

In continuous sign language recognition, traditional spatial feature extraction often relies on frame-by-frame static processing, which struggles to effectively capture dynamic changes between frames and limits the mining of key semantic information. Temporal modeling commonly employs fixed-scale convolutions, making it difficult to adapt to the variability of action boundaries, leading to blurred semantic segmentation and weakened temporal feature representation. To address these issues, this paper proposes a continuous sign language recognition network, Spatial-Temporal stage Network (S-TsNet). S-TsNet consists of two main stages: the spatial feature detection stage and the temporal feature detection stage. In the spatial feature extraction stage, a Mixed Spatiotemporal Block is designed, combining residual modules with an Element-level Multi-Scale Temporal Module (EMT) to model pixel-level inter-frame correlations and effectively capture dynamic information in key regions such as the hands. In the temporal modeling stage, a Multi-Scale Temporal Perception Module (MTP) is introduced, which employs hierarchical modeling and cross-scale fusion to adaptively model action boundaries. Experiments on the PHOENIX14 and PHOENIX14-T datasets demonstrate that S-TsNet significantly outperforms mainstream methods in recognition accuracy. Ablation studies and visualization further validate its effectiveness.

Object detection in images captured by unmanned aerial vehicles (UAVs) often presents challenges including the object is too small to extract effective features, the complex background is easy to produce a lot of noise interference. To address these challenges, we present a novel object detection algorithm, termed Adaptive Multi-Scale Feature Tower-YOLO (AMFT-YOLO). First of all, we propose a special detection head for detecting tiny objects called Task-decoupled Attention Head (TA-head), designed to mitigate the conflict between the requirements of localization and classification tasks on the feature maps. This allows for the efficient utilization of limited feature information. Secondly, we proposed the Bidirectional Multi-Scale Skip Fusion (BMSF) module, addresses the issue of missing information in tiny objects by aggregating global and local features. To solve the noise interference caused by complex backgrounds, we introduce the Dilated Convolutional Spatial Pyramid Pooling (DCSPP) module, which mitigates the effects of cluttered backgrounds through adaptive feature fusion. The effectiveness of our method is validated using the VisDrone and UAVDT datasets. Further ablation experiments with other methods confirm the robustness and adaptability of our method.

Deploying object detection models on resource-constrained embedded devices, such as mobile robots, requires lightweight and efficient network architectures. However, existing lightweight detectors often suffer from insufficient edge representation, redundant cross-scale features, and high inference costs, limiting their applicability. To address these challenges, this paper proposes LiteFuseNet, an efficient detection framework that enhances localization precision and reduces computational overhead. The network leverages multi-scale edge fusion to enhance edge awareness through edge response, and incorporates a hierarchical partial path aggregation with adaptive channel filtering feature fusion to enable efficient cross-layer guidance while reducing redundant information. In addition, it leverages a lightweight shared convolution and decoupled batch normalization head, sharing convolution across scales and using separate batch normalization for each scale to enhance efficiency. Extensive experiments on a self-constructed dataset and public benchmarks demonstrate that LiteFuseNet achieves superior detection performance with significantly reduced parameters and computation, and shows favorable results in mobile robot detection experiments, validating the effectiveness of the proposed network.

Tongue segmentation is an essential part for computer-aided tongue diagnosis. Since of similar color and texture between tongue body and non-tongue body, such as lips and face, existing methods produce the lack of accuracy and completeness for tongue segmentation results. Moreover, small samples in tongue datasets lead under-fitting on CNN-based methods which always produce poor segmentation. To solve these problems, we designed the quick accurate tongue segmentation net (QA-TSN) to segment tongue body. To alleviate small sample problem, in the proposed method, a tongue-style transfer generation net(T-STGN) was propose to synthesize tongue images. In T-STGN, a novel encoder-decoder structure with two encoder with a global rendering block was used to refine global characteristics of synthetic tongue images. For real-time tongue segmentation, quicker tongue segmentation net (QTSN) was proposed in QA-TSN. In QTSN, we used an encoder-decoder structure with modified partial convolution (MPConv) to expedite the computation for realtime segmentation. To smooth the segments of tongue body, a novel loss function of tongue segmentation loss (TSL) was proposed. In TSL, tongue edge loss (TEL) was used to smooth the boundary of segmentation of tongue body and tongue area loss (TAL) was proposed to improve the fragmentation of segmentation results. Experiments conducted on tongue datasets achieved an IoU of 98.0307 and a Dice score of 99.0738, with a frame rate of 75.35, outperforming all other methods involved in the experiment. These results demonstrate the effectiveness of the proposed QA-TSN.

Gesture recognition, an important method for Human-Drone Interaction (HDI), is often constrained by sensor limitations, such as sensitivity to lighting variations and field of view (FoV) restrictions. This letter proposes a real-time drone control system that integrates multimodal fusion gesture recognition with active perception control to overcome these challenges. We constructed a diversified arm gesture dataset and designed a lightweight, one-stage point cloud and image features fusion model, Adaptive Gate Fusion Network (AGFNet), for real-time inference on embedded devices. We applied motion compensation to mitigate delay errors caused by point clouds accumulation and network inference during movement, and fused the drone's velocity data and detection results using the Extended Kalman Filter (EKF) to enhance real-time performance. This enabled active perception through the optimization of the sensor's FoV using perception-aware Model Predictive Control (PAMPC). Experimental results demonstrate that the proposed model achieves a threefold improvement in inference speed compared to the baseline, reaching 97.93% mAP on the test set and outperforming single-sensor networks by approximately 5%. Real-world testing further confirms the system's applicability and effectiveness in indoor environments.