Search Results (324)

Pathogenic fungi and their synergistic interactions with bark beetles, leading to vascular dysfunction, physiological stress, and eventual tree mortality, increasingly threaten forest ecosystems. Because fungal colonization often precedes visible macroscopic symptoms, early detection remains a critical yet challenging task in forest health monitoring. This study proposes a real-time deep learning-based object detection framework for identifying harmful fungi in proximity to host trees to support early intervention strategies. A custom dataset comprising 8900 images was constructed to represent two classes: Healthy and Unhealthy trees, where fungal presence is detected either directly on the tree or within its immediate ecological vicinity (e.g., near root systems). A fine-tuned YOLOv11 detection architecture is developed and augmented with a squeeze-and-excitation (SE)-like attention mechanism to enhance texture-sensitive feature representation. The model is trained and evaluated using precision, recall, F1-score and mean Average Precision (mAP). Experimental results demonstrate an overall mAP@0.5 of 0.825, with class-wise average precision values of 0.926 (Healthy) and 0.724 (Unhealthy). The Healthy class achieved classification accuracy of 0.92, while 0.71 of Unhealthy instances were correctly detected. F1-Confidence and recall-Confidence metrics indicate that optimal operational performance occurs within a confidence threshold range of 0.30–0.35, balancing false positives and false negatives. Despite the approximately balanced class distribution (50.6% Healthy and 49.4% Unhealthy), detection performance for the Unhealthy class was comparatively lower because of its greater intra-class variability, heterogeneous fungal appearance, and subtle visual manifestations. Findings demonstrate the feasibility of deploying real-time object detection models for early-stage fungal surveillance and highlight the importance of confidence calibration for operational disease monitoring systems. Full article

Search-and-rescue (SAR) missions following earthquakes and other disasters require aerial video perception systems that do more than detect objects in isolated frames. Operators must maintain the identities of access points, vehicles, responders, hazards, and other mission-relevant targets despite UAV ego-motion, dust, debris, occlusion, scale variation, and abrupt scene transitions. This paper presents E’CHIT (Edge-Oriented Colour Histogram Instance-Guided Tracking), a deployment-oriented, operator-centric UAV tracking framework for real-world disaster-response applications. Its primary scientific contribution is an identity-stabilised, detector-assisted tracking methodology. YOLOv8-seg proposals trained on D’RespNeT initialise and refresh tracks; a Custom-RE3 recurrent module propagates target states through short detector dropouts; and a lightweight EOMC verifier, based on edge orientation, mean colour, and shape consistency, determines whether tracks should be accepted, refreshed, or reacquired. A scene-cut watchdog that combines luminance mean absolute difference (MAD) with HSV histogram divergence prevents stale identities from carrying over after hard edits or sudden feed changes. Custom-RE3 is the continuation module implemented and evaluated in this study. The surrounding E’CHIT wrapper follows an initialise–reseed–verify–reset cycle and is tracker-adaptable at the software-interface level: another compatible SOT or MOT continuation module can be integrated through adapter modifications, state and bounding-box conversion, and method-specific retuning, followed by independent validation. All reported quantitative results therefore apply to the Custom-RE3 implementation. D’RespNeT, the optional reinforcement learning (RL) warm start, the HUD, and the deployment stack support this central tracking contribution. D’RespNeT provides 28 polygon-annotated SAR classes. An author-developed PPO/SAC script is used only during offline detector training. In the reported runs, it produces different early optimisation trajectories for selected difficult or under-represented classes, while the default supervised schedule remains the strongest final global mAP reference. No RL policy runs during deployment; the detector architecture, parameter count, and inference graph remain unchanged. Evaluation on D’RespNeT and authentic disaster-response UAV footage shows that E’CHIT increases Success@IoU ≥ 0.5 from 0.62 to 0.79, reduces identity switches by approximately 71%, and maintains real-time 1080p performance, achieving 164–330 FPS for single-target tracking and 24–100+ FPS for end-to-end multi-target operation on an RTX-class GPU using FP16. The VOT2014, NT-VOT211, and VOTS2024 figures reproduce historical result spaces reported in the literature and include a clearly labelled, non-official E’CHIT operating-point marker solely for context. This marker was not produced using the corresponding official datasets, toolkits, reset rules, or submission routes; it is excluded from the primary quantitative claims and must not be interpreted as a leaderboard rank or a protocol-identical comparison. Overall, the system demonstrates how identity-stable UAV tracks can provide actionable operator cues for target monitoring, entry-point assessment, and UAV–UGV/ground-team coordination in cluttered disaster scenes. Full article

Multi-access edge computing (MEC) enables computation-intensive perception and decision-making tasks in vehicular networks to be offloaded to nearby edge servers. Existing approaches usually fix the artificial intelligence (AI) inference model, overlooking how model selection jointly affects latency, energy consumption, and service reliability. We propose a hierarchical model selection and control (HMSC) framework based on deep reinforcement learning (DRL) for MEC-assisted vehicular networks. The framework couples a vehicle-layer MAPPO component that provides a communication interface representation for subchannel assignment and energy accounting with a centralized MEC-layer soft actor-critic (SAC) agent that, under SDN orchestration, adaptively selects lightweight or high-fidelity AI models and allocates computational resources. Accordingly, the core contribution of this paper lies in MEC-side model-aware computation control under an explicitly defined subchannel-contention abstraction, rather than in physical-layer transmit-power optimization. Both layers are guided by a composite objective that integrates normalized end-to-end (E2E) latency, normalized energy consumption, and a deadline-violation penalty. Using a discrete-time simulation framework, HMSC reduces E2E latency compared with static inference and non-hierarchical DRL baselines and sustains a higher deadline satisfaction ratio (DSR) under constrained uplink throughput and varying traffic loads. The learned policy is load-aware, favoring high-fidelity inference under light load and lightweight inference under congestion; a post hoc analysis using YOLOv5-family accuracy reference further quantifies the inference-quality implications of this adaptive selection behavior. These results show that coordinated MEC-side control of AI model selection and computation, under a shared deadline-aware objective, provides a robust latency–energy trade-off for MEC-assisted vehicular networks. Full article

People counting is one of the key tasks in intelligent monitoring systems. However, accurately counting people in dynamic environments can be extremely challenging. This is especially true in mobile robot applications, where challenges such as a moving camera, varying conditions, and a limited field of view due to environmental obstacles arise. In such scenarios, the people counting task primarily involves visually detecting and grouping individuals to determine the total number of unique people. This paper presents a people counting algorithm based on visual people detection and clustering. The method utilizes the You Only Look Once (YOLO) detector to identify the bounding boxes of detected individuals and extract features from their corresponding regions of interest (ROIs). Additionally, the dimension of the extracted features is reduced using an encoder and clustered to distinguish individuals, with the number of clusters serving as an estimate of the number of people. The method was tested on a dataset containing 45 independent sequences with a total of 19,350 RGB images, complete with metadata for people detection and re-identification. This dataset encompasses various settings, particularly scenarios featuring mobile robots moving and capturing frames in indoor environments. The experiments demonstrate the effectiveness of the proposed method in different configurations. The best results were achieved using the YOLOv10n detector combined with K-means or SK-means clustering, yielding a Mean Absolute Error ($MAE$) of 1.11. The proposed encoder-based method significantly reduces clustering time and operates effectively within the limited resources available on mobile robotic platforms such as the Jetson Nano. Full article

Brain tumor detection in magnetic resonance images (MRI) is difficult because lesions often have weak boundaries, low contrast, and large variations in scale. Here, we present RD-YOLO, a YOLOv10n-based detector that redesigns neck-level feature fusion for lesion localization. It combines scale-normalized interpolation (SNI), enhanced group-shuffle convolution (GSConvE), and detail-preserving contextual fusion (DPCF). These modules regulate upsampled deep responses, enrich multi-receptive-field representation, and balance shallow structural and deep semantic features. On a public three-class brain tumor MRI detection dataset, RD-YOLO achieved Recall and F1-score values of $0.8326\pm 0.0144$ and $0.8441\pm 0.0058$. Its mAP@0.5 and mAP@0.5:0.95 values were $0.8858\pm 0.0048$ and $0.5490\pm 0.0076$, respectively. Among representative YOLO- and DETR-based baselines, the largest gain occurred for mAP@0.5:0.95, indicating more stable localization under stricter IoU evaluation. Full-factorial ablation showed that the complete configuration achieved the best F1-score and mAP@0.5 and maintained a balanced metric profile across the evaluated indicators. Full article

Accurate instance segmentation of recyclable waste is important for automated sorting and circular economy applications. Mixed metal and glass municipal waste is challenging because metal surfaces are often reflective, while glass objects have transparent boundaries and high visual variability. This paper presents a controlled computer vision framework that uses instance segmentation as an object-level perception layer for mixed metal and glass waste and links eligible detections to screening-level WARM-based environmental interpretation. A seven-class annotated dataset of 1667 images was evaluated using a fixed train/validation/test split. Model development used the validation split for architecture screening, optimization tuning, transfer learning analysis, seed selection, and NMS selection, and the held-out test set was used only for final evaluation. Using mask mAP@50:95 as the primary segmentation metric, the selected YOLOv8m-seg model achieved 0.9435 on the locked test set and obtained a higher mask mAP@50:95 than Mask R-CNN and RF-DETR-Seg under the same locked test protocol. Image-level bootstrap 95% confidence intervals were also used to characterize uncertainty around the locked test segmentation comparison. For environmental reporting, mass-bearing detections were mapped to class-specific mass priors and EPA WARM v16 factors rather than direct mask area-to-mass conversion; the masks were used for instance-level separation, localization, and visual verification, not as direct physical mass measurements. Under the stated assumptions, the framework produced screening-level environmental estimates of 1113.57 MJ of energy savings and 70.20 kg CO₂e of avoided emissions, decreasing to 445.67 MJ and 28.62 kg CO₂e after applying recycling rate priors. Overall, the framework provides a leakage-aware workflow linking instance segmentation to material-specific WARM-based environmental screening. Full article

Robust monocular distance estimation under varying illumination conditions is critical for autonomous driving safety. While state-of-the-art monocular 3D detection models achieve high accuracy in daylight conditions, they rely on computationally heavy architectures and degrade significantly in low-light environments. Lightweight 2D detectors (e.g., YOLO variants) offer real-time performance but lack the geometric constraints required for accurate depth estimation. To address this limitation, we propose the Anisotropic Geometry Loss (AGL) framework. This lightweight framework enforces ground-plane consistency through an anisotropic bottom-edge constraint derived from the pinhole camera model. In addition, a luminance-channel contrast enhancement module (CLAHE) is applied at inference to improve low-light visibility. Experimental results on the Dark-KITTI dataset show that the proposed method achieves an RMSE of 10.91 ± 0.68 m, improving over YOLOv10n (11.53 ± 0.56 m) and YOLOv26n (11.99 ± 0.58 m), while maintaining a 2.71 M-parameter footprint and real-time inference (>160 FPS). With CLAHE, RMSE is further reduced to 10.55 ± 0.72 m. Stratified by kinematic safety zone, the proposed method achieves 2.42 ± 0.03 m in the Near range (0–15 m), 5.94 ± 0.19 m in the Medium range (15–30 m), and 17.41 ± 1.25 m in the Far range (>30 m), corresponding to Euro NCAP AEB (Autonomous Emergency Braking) stopping distances. AGL provides its largest measurable accuracy improvement in the medium-distance range while maintaining comparable performance in the far-distance range. A complementary luminance-channel CLAHE preprocessor recovers bottom-edge gradients in synthetic and real low-light frames; zero-shot generalization is qualitatively corroborated on the ExDark dataset. These results demonstrate that explicit geometric constraints provide an effective and efficient solution for robust cross-illumination resistance in monocular distance estimation. The framework also shows practical potential for camera-only AEB systems deployed on edge-computing platforms aligned with Euro NCAP safety protocols. Full article

Vision-based estrus monitoring for goats has matured rapidly, yet endocrine preliminary feasibility evaluation in indigenous breeds is largely missing. Here we conducted a preliminary feasibility evaluation of whether behavioral estrus events captured by a machine vision pipeline align with reproductive endocrine dynamics in Chongming white goats. Ten multiparous does were monitored continuously for 21 days under standard housing conditions; daily jugular samples yielded 206 paired serum measurements of estradiol (E2), FSH, LH, and progesterone (P4). Behavioral events were extracted from infrared video by a YOLOv8-seg + MMPose pipeline and cross-checked against teaser-buck confirmation. Hormone profiles differed markedly between estrus and non-estrus windows (all p < 0.001): estradiol rose from 14.81 ± 4.17 to 21.79 ± 2.10 ng/L, and progesterone fell from 2386.70 ± 466.93 to 1719.68 ± 157.52 pmol/L. Estrus status correlated moderately with E2 (r = 0.459), FSH (r = 0.435), and LH (r = 0.446), and inversely with P4 (r = −0.406); the E2/P4 ratio produced the most pronounced per-event peak and is proposed as a preliminary sensitive auxiliary indicator. Vision-based detection reached 86.96% accuracy. The data indicate that camera-derived behavioral markers track real physiological state changes in this breed, supporting the deployment of vision-based estrus screening as a first-pass tool, with hormone assays reserved for confirmation in commercial Chongming white goat herds. Full article

The automation of fruit harvesting has become an important research topic in precision agriculture due to increasing labor shortages, rising production costs, and the need for improved harvesting efficiency. Among horticultural crops, strawberries present particular challenges for robotic harvesting because of their variability in size, shape, ripeness, and frequent occlusions caused by leaves and surrounding fruit. The objective of this work is to demonstrate the feasibility of a reproducible perception-to-manipulation framework for robotic strawberry harvesting based on commercially available hardware and established computer vision techniques, rather than to propose a novel object detection algorithm. The proposed system integrates a YOLOv3-based (You Only Look Once) object detector, monocular vision for fruit localization, and a Universal Robots UR5e collaborative manipulator. Strawberry coordinates estimated from monocular images are transformed into the robot reference frame and transmitted through the XML-RPC (Extensible Markup Language-Remote Procedure Call) protocol, enabling robot positioning. The system was experimentally validated in a controlled indoor environment under different artificial illumination conditions. The YOLOv3 detector achieved a ${\mathrm{mAP}}_{0.5:0.95}$ of 37.4%, a precision of 84.2%, a recall of 76.1%, and a latency of 6.5 ms per image (153.8 FPS). The experiments also demonstrated reliable communication between the perception and robotic manipulation modules, enabling the robotic arm to reach the estimated strawberry positions. The proposed framework provides a practical and low-cost solution for integrating deep-learning-based perception with robotic manipulation and establishes a solid basis for future work on localization accuracy, automated grasping, harvesting efficiency, and deployment in real agricultural environments. Full article

Real-time monitoring of endangered birds on edge AI hardware is constrained by a structural mismatch between modern lightweight detectors and mainstream NPU deployment toolchains. Recent attention-based detectors rely heavily on dynamic-shape operators that fall back to the host CPU on embedded NPUs, negating the advantages of lightweight architectures. To address this, we propose eBirdNet-Nano, a 1.05 M-parameter detector derived from YOLOv12n through a three-level NPU-friendly redesign: a static NPUConv block at the operator level, an NPU-C3k2 module together with an NPU-SE-Block at the module level, and a balanced 64-channel detection head at the head level. The resulting model achieves a 59% parameter reduction over YOLOv12n at only 5.8 GFLOPs while attaining an mAP@0.5 of 0.929 on a curated 24-species endangered-bird dataset collected in Macao. We further evaluate the model across four heterogeneous edge platforms—the Rockchip RK3588 (ARM + NPU), Kendryte K230 (RISC-V + KPU), Raspberry Pi 4B (pure ARM), and LicheePi 4A (pure RISC-V)—to characterize its behavior under distinct execution models. On the RK3588 NPU under INT8 quantization, eBirdNet-Nano delivers 13.83 ms inference latency and 26.76 ms end-to-end latency at 37.4 FPS, attaining the best parameter–latency balance and the highest parameter-normalized throughput (35.62 FPS/M) among six nano-scale YOLO variants, with an overall 3.53× end-to-end speedup over the YOLOv12n FP16 baseline that decomposes into a 2.97× architectural factor and a 1.19× quantization factor. Integrated into the EbirdEye field terminal, the same model sustains 23.5 ms thread-level end-to-end latency during live operation while supporting approximately 13.5 h of battery-powered runtime per charge. The proposed design offers a practical pathway toward deployable, low-power AI terminals for endangered-species conservation in resource-constrained field environments. Full article

X-ray inspection is crucial for pipeline weld non-destructive testing (NDT), yet automatic defect detection remains challenging due to low contrast, complex backgrounds, and significant variations in defect morphology. To address these issues, this paper proposes an improved YOLOv11-based method for X-ray weld images, integrating Smooth IoU (SIoU) bounding box regression and P2 shallow feature enhancement. First, to enhance the localization accuracy of elongated Region of Interest (ROI) targets in small-diameter pipe welds, the original CIoU loss is replaced with SIoU loss. By introducing an Angle Cost term, SIoU provides explicit directional constraints, guiding the predicted bounding box to align with the ground-truth orientation. Experimental results show the YOLOv11s + SIoU model achieves 99.5% mAP@50 and 99.9% precision, outperforming the baseline. Second, to improve the detection of larger defects (e.g., lack of fusion, incomplete penetration, and cracks) in long-distance pipeline welds, a P2 detection layer (stride 4) is added. This layer preserves high-resolution spatial details and shallow edge features that are typically lost during deep downsampling. Evaluated on a 960 × 960 input resolution, the YOLOv11s + P2 model achieves 93.07% precision, 94.8% mAP@50, and 72.01% mAP@50–95. The proposed method effectively combines directional constraint with shallow feature preservation, providing a robust solution for both ROI localization and large defect recognition in complex weld X-ray images. Full article

Breast ultrasound (BUS) is important for breast tumor detection, but speckle noise, low contrast, operator dependency, and limited medical datasets hinder robust deep learning. Although raw radiofrequency (RF) signals contain richer acoustic information than conventional B-mode images, multimodal fusion approaches often increase computational cost. To address these issues, this study proposes a curriculum transfer learning-based approach that sequentially exploits different ultrasound information representations during training. The proposed approach maintains a single detection model architecture rather than relying on complex multimodal input fusion. Phase, Envelope, and B-mode images generated from raw RF signals were defined as distinct input domains, and various training orders were evaluated. In addition, lightweight detection models based on YOLOv5, YOLOv8, YOLO11, and YOLO26 were compared to select the optimal model. A total of nine experimental settings, including single-modality training and curriculum learning conditions, were repeatedly evaluated using 100 random seeds. The experimental results showed that the proposed Phase-Envelope-B-mode (P-E-B) curriculum transfer learning strategy achieved the highest average mAP@50 among the evaluated training scenarios, with an approximately 2.08% relative improvement over single B-mode training under the fixed patient-level split and 100-seed repeated evaluation setting. The average convergence epoch was also lower than that of single B-mode training, indicating that the proposed strategy provided a favorable convergence profile while improving average validation performance. These results should be interpreted as proof-of-concept evidence obtained under a fixed patient-level split and 100-seed repeated evaluation setting, rather than as conclusive evidence of external clinical generalizability. Within this controlled small-data setting, the findings suggest that RF-derived representations may provide useful training-stage curriculum information for B-mode-based breast tumor detection while maintaining a single B-mode inference pathway. Full article

The worldwide spread of semiconductor devices has driven a surge in electronic waste (e-waste), which reached 62 million metric tons in 2022 and is projected to exceed 80 million metric tons by 2030. E-waste contains hazardous substances such as cadmium and mercury, yet also represents a $57 billion annual opportunity through the recovery of valuable and critical raw materials (CRMs). However, formal recycling rates remain stagnant at 22.3%, largely due to limitations of current automated sorting methods. These systems primarily rely on visible-light (RGB) imaging, which lacks the spectral resolution needed to distinguish chemically similar polymers, complex metal alloys, and composite substrates on printed circuit boards (PCBs). This paper presents a multidisciplinary synthesis of AI-driven detection and classification for e-waste, bridging materials science and computer vision through three interconnected themes. 1. Material and Economic Context: The toxicological risks and economic drivers of semiconductor recycling are characterized, framing fine-grained material identification as essential for a circular economy. 2. Multispectral Sensing & Fusion: Sensing modalities such as near-infrared (NIR), hyperspectral imaging (HSI), and X-ray fluorescence (XRF) are assessed, and sensor fusion strategies, including early, late, and intermediate fusion, are reviewed for high-throughput industrial settings. 3. Deep Learning Benchmarking: 11 publicly available PCB datasets are analyzed, and the YOLO series (YOLOv3–YOLOv12) is compared with leading non-YOLO detectors, including Faster R-CNN, RT-DETR-L, and RetinaNet. The results show that while YOLOv9s achieves a peak mAP@0.5 of 56.5% and YOLOv11s offers an optimal industrial profile (37.2% mAP@0.5:0.95 at 115 ms edge inference), all RGB-based models fail to detect visually ambiguous surface-mount devices (SMDs), with mAP values below 12%. This confirms a performance ceiling for purely visual systems. The review concludes that transitioning from RGB-centric to multispectral fusion architectures is the primary research frontier and proposes a roadmap for standardized multimodal datasets and edge-deployable fusion models to enable next-generation, high-recovery automated recycling. Full article

Background/Objectives: Spontaneous intracranial hypotension (SIH) is caused by spinal cerebrospinal fluid (CSF) leakage and is typically diagnosed by clinical presentation and characteristic MRI signs; however, objective tools for monitoring physiological changes and treatment response remain limited. Cine phase-contrast MRI (PC-MRI) enables noninvasive quantification of aqueductal CSF dynamics, yet reliable analysis is challenging since the cerebral aqueduct is extremely small and susceptible to low contrast, partial volume effects, and ROI-dependent measurement variability—particularly in SIH where CSF pulsatility is often reduced. Methods: We propose an end-to-end automated framework that integrates (1) a cascade localization–segmentation strategy, consisting of Tiny YOLOv4 detection followed by MultiResUNet segmentation on a YOLOv4-derived cropped ROI; (2) physiology-informed pulsatility-based segmentation (PUBS) to refine anatomical masks into functional flow ROIs; and (3) one-dimensional convolutional neural networks (1D-CNNs) to extract exploratory waveform morphology features from 32-phase cardiac-cycle velocity waveforms. The study includes 39 participants, yielding 59 cine PC-MRI examinations: 11 controls, 28 Pre-treatment SIH scans and 20 Post-treatment Recovery scans. Results: The cascade model significantly improves segmentation robustness compared with a full-image baseline, achieving higher Dice scores and markedly lower boundary errors across cohorts (e.g., Pre-treatment SIH HD95: 1.66 ± 0.74 px vs. 15.37 ± 44.98 px). PUBS refinement reduces quantification deviation from expert manual references in SIH (mean relative error: 7.4% to 5.6%) and improves diagnostic performance for multiple hemodynamic parameters (e.g., downward mean flow AUC: 0.747 to 0.792). For waveform morphology analysis, the end-to-end 1D-CNN classifier was evaluated using repeated-seed participant-level grouped LOOCV. The repeated-seed ensemble prediction showed modest out-of-sample discrimination between Normal controls and Pre-treatment SIH scans, with an AUC of 0.646, a bootstrap 95% confidence interval of 0.455–0.826, and a permutation-test p-value of 0.072. Separately, exploratory analysis of the final baseline-trained 1D-CNN latent space showed marked, apparent Normal-versus-SIH separability and an intermediate recovery distribution in PCA space, suggesting that aqueductal waveform morphology may encode SIH-related physiological information. Conclusions: These findings suggest that SIH-related information may be reflected not only in flow magnitude but also in aqueductal CSF waveform morphology. However, the modest and statistically non-significant out-of-sample performance of the end-to-end 1D-CNN classifier indicates that morphology-based AI features should currently be regarded as exploratory biomarker candidates rather than validated stand-alone diagnostic tools. Larger independent cohorts are required to confirm their reproducibility, physiological meaning, and clinical utility. Full article

Object detection in complex industrial environments is prone to being affected by insufficient dynamic weighting of local and global features, as well as illumination variations and impurities. Moreover, existing models suffer from excessive model complexity, which directly impairs computational efficiency. To more accurately distinguish Chinese herbal materials with diverse morphologies, this paper proposes the MobileAttn module. Drawing on the idea of token representation in the Transformer architecture, this module extracts contextual information through global feature compression, fuses it with tokens to generate a spatial attention map, and realizes dynamic recalibration of convolutional features. This process enhances the feature weights of key semantic regions, suppresses redundant background information, and improves feature discriminability. To address illumination interference, brightness-aware weights are combined with dual-path (channel and spatial) attention for global control, dynamically reducing the impact of illumination; this component is named LightAttn. When Chinese herbal materials contain common industrial unknown impurities (e.g., small stones and weeds), an impurity detection auxiliary module, a post-processing step independent of the main detection network, is proposed. This module refines Non-Maximum Suppression (NMS) logic to distinguish target Chinese herbal materials from interfering impurities. Subsequently, it accurately locates and marks impurities on the conveyor belt, thereby achieving effective unknown impurity detection. Experimental results demonstrate that, compared with the original YOLOv11 on the Chinese herbal materials detection task, the optimized model achieves a 1.7% improvement in the overall mean Average Precision (mAP@0.5:0.95). On a per-class basis, gains are particularly pronounced for certain challenging high-aspect-ratio Chinese herbal materials. Prunella vulgaris and orange peel achieve respective AP improvements of 5.8% and 4.1%. Meanwhile, the model parameter count is reduced by 23.1% and the computational complexity by 20.3%. The F1-Score of the impurity detection results is 86.38%, verifying the effectiveness of the impurity detection auxiliary module. Full article