Real-time automated steel surface inspection on the production line is challenging. Defect masks must be produced at the camera’s acquisition rate and within the power envelope of an embedded GPU. This paper evaluates two contrasting YOLO architectures, the attention-centric YOLOv12 and the convolution-centric YOLOv26, for instance segmentation of steel surface defects on three platforms: an RTX 5090 workstation, a Jetson AGX Orin, and a Jetson Orin Nano. Twenty-four models were trained in a 2×2 ablation over weight initialization (scratch, pretrained) and class sampling (standard, balanced) on a real industrial dataset of 3,097 images and 13,318 instances across 8 classes (11-fold imbalance). The models were then evaluated under PyTorch and under TensorRT in 16-bit floating-point (FP16) and 8-bit integer (INT8) precision. Accuracy is reported on a leak-free test subset after a split-integrity audit. The best configuration, YOLOv26 pretrained with FP16, achieves 0.43 mean average precision (mAP50-95) at large scale. All YOLOv26 pretrained models surpass the 30 frames-per-second (FPS) real-time budget on both Jetson devices under FP16: 89.9 FPS at nano on the Orin Nano and 146.7 FPS on the AGX Orin. YOLOv12 also deploys under TensorRT FP16 on all three platforms and, trained from scratch, matches YOLOv26 pretrained accuracy at the medium scale. The training-time choices (architecture, initialization, and sampling) are statistically indistinguishable; therefore, deployment is the deciding factor. FP16 preserves accuracy within 0.02 mAP50-95. However, INT8 segmentation engines fail to build on both Jetson devices under TensorRT 10.3 and their accuracy is unreliable on the workstation.
To achieve an improved balance between detection accuracy and computational efficiency for unmanned aerial vehicle (UAV)-based small-object vehicle detection, this paper proposes LHCANet, a lightweight detection model for resource-constrained scenarios. The core problem addressed in this work is how to reduce redundant detection structures while preserving effective small-object representation under scale variation, directional changes, and complex background interference. LHCANet is optimized through three corresponding strategies. First, a task-oriented branch pruning strategy is adopted to remove the redundant large-scale branch. Second, a C2f-directional multi-scale spatial enhancement (C2f-DMSE) module is introduced to enrich spatial structural representation across different scales and directions. Third, we design a bidirectional spatial–channel aggregation (BSCA) module to mitigate background interference, and a bidirectional semantic–detail interaction (BSDI) module to enhance target-relevant features prior to detection. Compared with the YOLOv8n baseline, LHCANet reduces the number of parameters from 3.1M to 2.02M (a 34.8
Real-time object detection for all-day UAV perception remains challenging because weak small objects occupy only a few pixels and are easily suppressed by complex backgrounds. While RGB-infrared fusion can improve robustness under illumination variations, nighttime imaging, and weak-texture conditions, existing enhancement methods often rely on global statistics, potentially amplifying salient backgrounds rather than true targets. Moreover, dense cross-modal interaction over the entire feature map introduces redundant computation and may propagate shadows, background clutter, and infrared noise. To address these issues, this paper proposes TDSF-Net, a target-driven sparse fusion network for real-time RGB-infrared weak small object detection. TDSF-Net follows an enhance-then-fuse paradigm. During the enhancement stage, TALEN generates target-aware guidance maps based on dual-modal statistical responses and cross-modal differences, concentrating computation on potential small object regions and avoiding excessive background processing. In the fusion stage, SGIF performs cross-modal interaction only within high-value candidate regions, while employing global stable fusion and gated recomposition to integrate features reliably, thereby reducing unnecessary computation and preventing propagation of irrelevant background information. Experiments on DroneVehicle, VEDAI, and M3FD demonstrate the effectiveness and generalization capability of TDSF-Net. On the UAV-based DroneVehicle dataset, TDSF-Net achieves state-of-the-art performance with 86.4
Underwater sonar image object detection is challenged by low contrast, strong speckle noise, and the underutilization of acoustic shadow information. To address these issues, we propose SonarOBB, an acoustic-structure-guided oriented object detection framework specifically designed for sonar imaging. First, a Contrast and Response Joint Enhancement (CRJE) module is introduced to improve feature discriminability by combining adaptive contrast normalization with global response modulation. Second, an Acoustic Shadow Aware Module (ASAM) enhances shadow-aware feature representation through multi-directional spatial attention, enabling more effective exploitation of structural cues in sonar images. Furthermore, an oriented bounding box regression strategy is adopted to accurately capture arbitrarily oriented and elongated underwater objects. Extensive experiments on four sonar datasets demonstrate consistent improvements over strong OBB baselines. Specifically, our method achieves improvements in mAP50 from 0.7534 to 0.8009 on SIMD and from 0.9482 to 0.9952 on SCTD. The proposed framework maintains real-time inference speed (98 FPS) with low computational overhead, demonstrating a favorable accuracy–efficiency trade-off under challenging low-contrast and noisy sonar conditions. The code is released at https://github.com/wangxu777777/SonarOBB .
Small unmanned aerial vehicles (UAVs) are difficult to detect in complex environments because of low pixel occupancy and strong background interference. This paper proposes ATU-Net, a YOLO11-based detector that integrates an enhanced multi-scale aggregation (EMA) backbone, a shuffle-fusion cross stage partial (SFCSP) neck module, a feature-sharing (FS) detection head, and Focaler-ShapeIoU. EMA combines progressive multi-scale feature aggregation (PMFA) for progressive detail–context aggregation with oriented asymmetric convolution (OAConv) for directional structural enhancement. SFCSP uses lightweight feature extraction and partial cross-stage fusion to reduce redundant computation, while the FS head shares feature transformations and prediction layers across scales and introduces multi-directional difference convolution to strengthen weak edge and texture-gradient responses. Focaler-ShapeIoU combines overlap-quality focusing with shape-aware geometric constraints. On DUT Anti-UAV and Det-Fly, ATU-Net improves YOLO11n by 2.7 and 3.1 percentage points in mAP50 and by 3.6 percentage points in mAP50–95 on both datasets. On VisDrone2019, the gains are 2.0 and 1.6 percentage points, respectively. Deployment on a Jetson Orin NX reaches 38.0 FPS under the tested 25 W configuration, showing that the accuracy gain is retained with real-time embedded inference despite additional runtime overhead.
Driven by the rapid expansion of the low-altitude economy, unmanned aerial vehicles (UAVs) have emerged as pivotal platforms for aerial perception. However, balancing detection accuracy and computational efficiency for micro-targets remains a formidable challenge. To address this, we propose LSO-YOLO, a novel real-time detector optimized for UAV-based small-object detection. Our primary contributions are threefold: first, a multi-scale dilated context aggregation (MDCA) module is designed to replace the final down-sampling convolution in the backbone. By employing sequential dilated depth-wise convolutions and a lightweight channel attention mechanism, the MDCA expands the receptive field and suppresses background noise, effectively mitigating the loss of fine-grained details caused by aggressive down-sampling; second, the C2f module in the neck is redesigned by substituting the second convolution in the bottleneck with parallel grouped convolutions, which enhances feature diversity and multi-scale representation while reducing computational overhead; finally, we optimize the overall architecture and introduce a lightweight down-sampling strategy to mitigate inter-channel interference. Experimental results on the VisDrone2019 dataset demonstrate that the proposed method improves mAP@0.5 and mAP@0.5:0.95 by 2.9 https://github.com/diapfer/LSOYOLO .
Reliable inspection of transmission line defects is critical for the safe and stable operation of modern power grids. Although UAV-based visual inspection combined with deep learning has been widely adopted, existing methods still suffer from high miss rates and limited accuracy when handling small-scale components and complex backgrounds. To address these limitations, this paper proposes TLD-YOLO, an enhanced object detection framework built upon the YOLOv5 baseline for fast and accurate transmission line defect inspection, which incorporates four key architectural improvements. First, a parameter-free Simple Attention Module (SimAM) is embedded into the backbone to strengthen multi-scale feature representation and improve sensitivity to small objects. Second, the Spatial–Channel Enhancement (SCE) module employs scale–channel interaction gating before multi-scale concatenation. This mechanism suppresses redundant scale responses and enhances discriminative feature fusion in dense-target and large-scale-variation scenarios. Third, the conventional neck network is replaced by a weighted Bi-directional Feature Pyramid Network (BiFPN) augmented with the Convolutional Block Attention Module (CBAM), enabling more effective cross-scale information aggregation while mitigating feature loss. Finally, a decoupled detection head separates classification from box regression and objectness prediction. The task-specific branches reduce task interference and improve detection accuracy. Experimental results on the self-constructed transmission line defect (TLD) dataset show that TLD-YOLO improves precision, recall, and mAP50 over YOLOv5 by 5.1
Small object detection in unmanned aerial vehicle (UAV) imagery holds significant application value across diverse domains, including urban security surveillance, agricultural crop monitoring, and disaster emergency response. However, existing methodologies encounter a fundamental trade-off dilemma between detection accuracy and computational efficiency, primarily attributed to extremely small target scales, complex background interference, and stringent computational resource constraints. To address these challenges, we propose AEI-DETR, which achieves superior detection accuracy compared to existing methods while substantially reducing model complexity through systematic architectural innovations. Specifically, we first design an adaptive spatial cross-stage backbone (ASCB) that reduces the complexity of global dependency modeling from quadratic O(L2) with respect to sequence length to linear O(L) through state space modeling, incorporating cross-stage feature aggregation and spatial coordinate enhancement mechanisms to enable efficient lightweight feature extraction. Subsequently, we develop an efficient lightweight feature neck (ELFN) that optimizes cross-scale feature fusion through adaptive fusion strategies within multi-feature modulation modules, leveraging depthwise separable convolutions to dramatically reduce computational overhead. Furthermore, we employ a bounding box regression strategy that combines internal constraints with distance optimization, constructing hierarchical optimization pathways through auxiliary bounding box mechanisms and point distance constraints to provide more stable and precise supervisory signals for small target localization. Experimental results on public datasets demonstrate that AEI-DETR significantly outperforms existing state-of-the-art methods.
Weeds in complex field environments are often small, densely distributed, and affected by severe background clutter, posing substantial challenges to the development of high-accuracy and low-complexity detection models. To balance detection accuracy with the requirements of lightweight deployment, this paper proposes a YOLOv11-based lightweight weed detection model, termed multi-level feature aggregation-based YOLO (MFA-YOLO). First, a gated progressive feature fusion (GPFF) module is proposed to enhance hierarchical feature interaction via learnable weighting and channel-wise gating. Second, an efficient feature aggregation (EFA) module strengthens fine-grained texture and semantic representation through lightweight attention-guided aggregation. Third, a P2-enhanced bidirectional attention feature pyramid network (P2-BAFPN) exploits shallow high-resolution features and bidirectional cross-scale fusion for small-scale weed perception. Finally, a lightweight shared detail-enhanced convolutional detection head (LSDECD) reduces head redundancy while improving detail-sensitive classification and localization in cluttered backgrounds. On the Lincoln Beet dataset, MFA-YOLO reduces the number of parameters, GFLOPs, and model size by 47.66
The intersection of adaptive machine learning and real-time systems presents one of the most pressing engineering challenges in modern computer vision. Online Evolutive Learning (OEL) enables models to improve continuously through environmental interaction during deployment. However, the computational overhead inherent to these adaptive mechanisms conflicts directly with the microsecond-to-millisecond latency budgets of real-time image processing pipelines. This survey provides a systematic review of the OEL landscape from the perspective of real-time image processing, which is organized along two principal axes: (i) the key algorithmic and hardware technologies that constitute the OEL—spanning continual learning, test-time adaptation, few-shot and meta-learning, self-supervised evolution, model compression, hardware-software co-design, generative AI and large vision-language model integration, and edge-cloud collaboration—and (ii) the downstream real-time image processing tasks to which OEL has been applied. For each technology and application area, we survey the state of the art, identify the fundamental conflicts between adaptability and real-time performance, and discuss algorithmic and architectural strategies to reconcile them. We conclude by proposing a unified evaluation protocol and identifying open challenges, including convergence guarantees under bounded latency, safe deployment in safety-critical systems, and the standardisation of real-time OEL benchmarks.
Recent years have seen growing interest in deploying computer vision (CV) applications at the network edge to meet real-time, bandwidth, and security constraints. This has increased the need to run multiple CV pipelines concurrently on resource- and power-constrained edge GPU accelerators, where shared resources, DVFS, and thermal throttling introduce complex interference effects that existing single-workload models do not capture. Benchmarking-based approaches are costly and do not scale across diverse combinations of workloads and platforms. This paper proposes a mathematical model to predict throughput and energy consumption for concurrently executing CV workloads on edge GPU accelerators. The model accounts for shared resource contention, host–device communication, DVFS, and thermal effects, and can be integrated into functional simulation frameworks for edge–cloud deployment studies. Validation across multiple edge platforms and representative CV workloads shows that the model predicts throughput and power consumption with an average error below 10%. The approach remains accurate under different compilation optimizations and precision levels, while requiring only minimal baseline benchmarking.
Face detection serves as a critical first step in many computer vision applications and has advanced significantly with the advent of convolutional neural networks (CNNs). Despite their effectiveness, CNN-based algorithms often require substantial computational resources, making hardware acceleration using FPGAs essential for real-time processing. FaceBoxes, one of the most accurate face detection algorithms, poses particular challenges for FPGA implementation due to its high computational complexity. In this paper, we propose an FPGA-based CNN inspired by FaceBoxes that significantly reduces hardware resource utilization and memory access requirements. The proposed network increases the number of layers while reducing filter sizes, and an optimized hardware architecture tailored to this network is presented. The design employs only 3 × 3 convolutional kernels, achieving a minimal accuracy degradation of 2.8
Accurate segmentation of the bolus and hyoid bone is essential for the automated detection of dysphagia in videofluoroscopic swallowing studies (VFSS). To address the challenges of blurred boundaries and substantial scale differences between the two target structures in VFSS images, this study considers feature extraction, multi-scale fusion, and branch interaction, and proposes a lightweight segmentation network, termed MSSA-Net, based on Mobile U-ViT. An edge detection module (ED) is introduced to enhance boundary features and efficiently extract critical information. During the downsampling process, a multi-scale self-attention mechanism (MSSA) is employed to capture dependencies among different spatial locations in the feature maps. In addition, a lightweight convolution—Transformer fusion module is designed to suppress redundant responses and integrate local details with global semantic information. Experimental results demonstrate that the proposed method can accurately and simultaneously track the bolus and hyoid bone. MSSA-Net-L achieves an mIoU of 83.64
The vision-based detections using deep learning need to localize and recognize the object of interests on continuous image sequence/video stream at low latency. Because of the intrinsic limitation of feature discriminability of non-salient objects inside images and over-complicated structures of present models, there is much space for improvement regarding detection efficiency and accuracy for such kind of targets. To address these problems, we propose a refined detection model named CDFI-YOLO. Our main contributions include: first, we introduce a Tri-Directional Attention (TDA) module into the backbone of the YOLOv11 baseline framework, modeling the spatial position information via a three-way attention mechanism to capture richer fine-grained features; second, we devise a Cross-Dimension Feature Interaction (CDFI) module to further refine the feature exchange procedure within our model, enhancing the recognition accuracy for non-salient objects; finally, we adopt a Lightweight Coupled Detection Head (LCDH), greatly reducing the parameter count of the baseline detection head thanks to its parameter-sharing GSConv structure, resulting in a 27.6
Low-altitude UAV obstacle detection is challenged by small obstacles that are easily disturbed by complex backgrounds, illumination changes, and motion blur, as well as by the strict computational constraints of onboard platforms. Although existing detection models can achieve high accuracy, they often suffer from large parameter sizes and redundant computations, making real-time edge device deployment difficult; meanwhile, direct lightweighting may weaken small-obstacle feature representation and lead to accuracy degradation. To address these issues, this paper designs a lightweight bidirectional extended path aggregation network (LBPAN) to enhance cross-scale feature fusion and introduces a balanced attention module (BAM) to improve small-object feature extraction. Meanwhile, a lightweight-C3k2 (L-C3k2) is proposed to reduce network redundancy. An improved Slim pruning method and a joint knowledge distillation strategy are adopted to achieve efficient model compression and accuracy compensation. On this basis, taking YOLOv11n as the baseline, a lightweight detection model lightweight bidirectional fusion YOLO (LBF-YOLO) is constructed and validated on a self-built low-altitude obstacle detection dataset. Experimental results show that, under significant lightweighting (parameters reduced from 2.6 to 1.4 M, FLOPs decreased from 6.6 to 4.4 G), mAP_0.5 increases from 85.3 to 88.4 mAP_0.5:0.95 rises from 62.8 to 66.4
Accurate safety-helmet detection in underground coal mines remains a critical challenge due to extreme environmental factors and the prevalence of small-scale, partially occluded targets. To address these bottlenecks, this paper proposes LCV-YOLO, a lightweight and robust detection framework specifically architected for complex mining environments. We re-engineer the backbone via a tailored LC-C2f-LSK module to dynamically expand the receptive field for minute targets, and introduce a CGA Fusion mechanism at the neck to suppress mining noise while adaptively enhancing target semantics. Furthermore, an efficiency-driven VoV-GSCSP module is implemented to prioritize feature reuse under strict computational constraints. Extensive experiments demonstrate that LCV-YOLO outperforms the YOLOv8 baseline with a 5.7
Lane line detection, object detection, and drivable area segmentation serve as the core tasks of autonomous driving environment perception. The accuracy and real-time performance of this perception directly impact the operational safety of intelligent transportation systems. Existing multi-task vision models typically employ a parallel branch structure, resulting in inadequate information sharing and guidance among tasks. In occluded or blurred areas, the accuracy is relatively low, which can further give rise to the risk of traffic accidents and pose a potential threat to public safety. Additionally, it is challenging to achieve a balance between detection accuracy and inference latency, which may also induce safety risks. To tackle these issues, this paper presents a lightweight multi-task learning framework for safety-critical autonomous driving perception. This framework integrates the three aforementioned tasks into a unified model to concurrently enhance real-time processing efficiency and perception reliability. First, Cross-Task Attention Fusion (CTAF) is proposed. Through the Cross-Task Attention Block (CTAB), the collaborative modeling capability of lane line detection and drivable area segmentation is strengthened. Second, Dynamic Path Fusion with Transformer (DPFT) is introduced to dynamically generate the fusion weights of feature maps, thereby enhancing the model’s adaptability to different driving scenarios. Finally, a semantic boundary detection auxiliary task is introduced to improve edge perception and the localization accuracy of lane lines and road boundaries. The overall network employs a shared encoder-task-specific decoder structure and utilizes ELAN-Net as the backbone network for feature extraction. Experimental results show that the proposed method achieves an inference speed of 78 FPS while maintaining a lightweight design and improving multi-task perception accuracy. It strikes a balance between accuracy and real-time performance in multi-task perception, thereby supporting safety-critical perception by reducing perception latency and improving perception reliability.
Timely and accurate detection of steel surface defects is crucial for ensuring industrial product quality and production safety. However, existing methods often struggle to balance detection accuracy, inference speed, and model lightweightness, especially under complex industrial conditions with multi-scale defects. To address this issue, this work proposes EG-YOLO, a lightweight and efficient detection framework based on YOLO11n for real-time steel surface defect inspection. Specifically, the Efficient Multi-Scale Attention (EMA) module is embedded into the backbone to enhance multi-scale feature representation through cross-spatial learning, while Generalized Intersection over Union (GIoU) loss is adopted to improve bounding box regression and localization of small and irregular defects. Experiments on the NEU-DET dataset show that EG-YOLO achieves 79.7
Semantic segmentation of point clouds is a fundamental task in environmental perception for autonomous driving. Due to the irregularity and sparsity of LiDAR data, efficiently processing large-scale unstructured point clouds remains a significant challenge. This paper presents LERPNet, a lightweight range-view-based semantic segmentation network. Specifically, we replace the Stem layer in FRNet with the proposed FD-Stem, which improves feature representation while reducing computational complexity. The proposed Adaptive Spatial-Channel Attention Enhancement module introduces long-range modeling capability with limited additional parameters, enabling effective learning of both spatial and channel-wise representations. Furthermore, the Range Point Nearest Neighbor Fusion Head alleviates the many-to-one mapping issue with lower inference latency, thereby improving segmentation performance. Extensive experiments demonstrate that LERPNet achieves competitive mIoU of 70.1
To address the issue of low detection accuracy for hazardous items caused by complex backgrounds, severe occlusion, multi-scale variations, and diverse poses in X-ray security inspection images, this study adopts the YOLOv12 network as the baseline model. The ADown module is introduced to improve downsampling efficiency while maintaining lightweight performance. Additionally, the BN-CGLU (batch normalization convolutional gated linear unit) is incorporated into the A2C2f module to enhance the model’s nonlinear representation capability. Furthermore, the SEAM (separated and enhancement attention module) attention mechanism is integrated into the detection head to improve detection precision under occlusion conditions. Experiments conducted on the public SIXray dataset demonstrate that, compared to YOLOv12, the proposed method reduces computational complexity by 15.5
