Unified Spatial-Frequency Modeling and Alignment for Multi-Scale Small Object Detection

Symmetry, Journal Year: 2025, Volume and Issue: 17(2), P. 242 - 242

Published: Feb. 6, 2025

Small object detection in aerial imagery remains challenging due to sparse feature representation, limited spatial resolution, and complex background interference. Current deep learning approaches enhance performance through multi-scale fusion, leveraging convolutional operations expand the receptive field or self-attention mechanisms for global context modeling. However, these methods primarily rely on spatial-domain features, while introduces high computational costs, conventional fusion strategies (e.g., concatenation addition) often result weak correlation boundary misalignment. To address challenges, we propose a unified spatial-frequency modeling alignment framework, termed USF-DETR, small detection. The framework comprises three key modules: Spatial-Frequency Interaction Backbone (SFIB), Dual Alignment Balance Fusion FPN (DABF-FPN), Efficient Attention-AIFI (EA-AIFI). SFIB integrates Scharr operator edge detail extraction FFT/IFFT capturing frequency-domain patterns, achieving balanced of semantics local details. DABF-FPN employs bidirectional geometric adaptive attention significance expression target area, suppress noise, improve asymmetry across scales. EA-AIFI streamlines Transformer mechanism by removing key-value interactions encoding query relationships via linear projections, significantly boosting inference speed contextual Experiments VisDrone TinyPerson datasets demonstrate effectiveness improvements 2.3% 1.4% mAP over baselines, respectively, balancing accuracy efficiency. outperforms state-of-the-art

Language: Английский

---

Integration of YOLOv8 Small and MobileNet V3 Large for Efficient Bird Detection and Classification on Mobile Devices

AI, Journal Year: 2025, Volume and Issue: 6(3), P. 57 - 57

Published: March 13, 2025

Background: Bird species identification and classification are crucial for biodiversity research, conservation initiatives, ecological monitoring. However, conventional techniques used by biologists time-consuming susceptible to human error. The integration of deep learning models offers a promising alternative automate enhance recognition processes. Methods: This study explores the use bird in city Zacatecas. Specifically, we implement YOLOv8 Small real-time detection MobileNet V3 classification. were trained tested on dataset comprising five species: Vermilion Flycatcher, Pine Mexican Chickadee, Arizona Woodpecker, Striped Sparrow. evaluation metrics included precision, recall, computational efficiency. Results: findings demonstrate that both achieve high accuracy identification. excels detection, making it suitable dynamic monitoring scenarios, while provides lightweight yet efficient solution. These results highlight potential artificial intelligence ornithological research improving reducing manual efforts.

Language: Английский

---

An improved YOLOv11 algorithm for small object detection in UAV images

Signal Image and Video Processing, Journal Year: 2025, Volume and Issue: 19(6)

Published: May 2, 2025

Language: Английский

---

AMFE‐YOLO: A Small Object Detection Model for Drone Images

IET Image Processing, Journal Year: 2025, Volume and Issue: 19(1)

Published: Jan. 1, 2025

ABSTRACT Drones, due to their high efficiency and flexibility, have been widely applied. However, small objects captured by drones are easily affected various conditions, resulting in suboptimal surveying performance. While the YOLO series has achieved significant success detecting large targets, it still faces challenges target detection. To address this, we propose an innovative model, AMFE‐YOLO, aimed at overcoming bottlenecks Firstly, introduce AMFE module focus on occluded thereby improving detection capabilities complex environments. Secondly, design SFSM merge shallow spatial information from input features with deep semantic obtained neck, enhancing representation ability of reducing noise. Additionally, implement a novel strategy that introduces auxiliary head identify very targets. Finally, reconfigured head, effectively addressing issue false positives small‐object precision object AMFE‐YOLO outperforms methods like YOLOv10 YOLOv11 terms mAP VisDrone2019 public dataset. Compared original YOLOv8s, average improved 5.5%, while model parameter size was reduced 0.7 M.

Language: Английский

---

Improving Object Detection Accuracy in Football Matches Using ECABiF-Y5n

Computer Science and Application, Journal Year: 2025, Volume and Issue: 15(02), P. 83 - 93

Published: Jan. 1, 2025

Language: Английский

---

Mitigating In-Transit Vision Noise for Enhanced Vehicle Safety

Published: May 4, 2025

Language: Английский

---

YOLO-LPSS: A Lightweight and Precise Detection Model for Small Sea Ships

Journal of Marine Science and Engineering, Journal Year: 2025, Volume and Issue: 13(5), P. 925 - 925

Published: May 8, 2025

The accurate detection of small ships based on images or vision is critical for many scenarios, like maritime surveillance, port security, and navigation safety. However, achieving a challenge cost-efficiency models; while the models could meet this requirement, they have unacceptable computation costs real-time surveillance. We propose YOLO-LPSS, novel model designed to significantly improve ship accuracy with low cost. characteristics YOLO-LPSS are as follows: (1) Strengthening backbone’s ability extract emphasize features relevant objects, particularly in semantic-rich layers. (2) A sophisticated, learnable method up-sampling processes employed, taking into account both deep image information semantic information. (3) Introducing post-processing mechanism final output resampling process restore missing local region high-resolution feature map capture global-dependence features. experimental results show that outperforms known YOLOv8 nano baseline other works, number parameters increases by only 0.33 M compared original YOLOv8n 0.796 0.831 AP50:95 classes consisting mainly targets (the bounding box target area less than 5% resolution), which 3–5% higher vanilla recent SOTA models.

Language: Английский

---

YOLOv8-SAMURAI: A Hybrid Tracking Framework for Ladder Worker Safety Monitoring in Occlusion Scenarios

Buildings, Journal Year: 2025, Volume and Issue: 15(11), P. 1836 - 1836

Published: May 27, 2025

Monitoring worker safety during ladder operations at construction sites is challenging due to occlusion, where workers are partially or fully obscured by objects other workers, and overlapping, which makes individual tracking difficult. Traditional object detection models, such as YOLOv8, struggle maintain continuity under these conditions. To address this, we propose an integrated framework combining YOLOv8 for initial the SAMURAI algorithm enhanced occlusion handling. The system was evaluated across four scenarios: non-occlusion, minor major multiple overlap. results indicate that, while performs well in non-occluded conditions, accuracy declines significantly severe occlusions. integration of improves stability, identity preservation, robustness against occlusion. In particular, achieved a success rate 94.8% 91.2% overlap scenarios—substantially outperforming alone maintaining continuity. This study demonstrates that YOLOv8-SAMURAI provides reliable solution real-time monitoring complex environments, offering foundation improved compliance risk mitigation.

Language: Английский

---

SO-RTDETR for Small Object Detection in Aerial Images

Published: Oct. 25, 2024

In aerial image object detection, small targets present significant challenges due to limited pixel information, complex backgrounds, and sensitivity bounding box perturbations. To tackle these issues, we propose SO-RTDETR for detection. The model introduces a Cross-Scale Feature Fusion with S2 (S2-CCFF) module, Parallelized Patch-Aware attention (PPA) the Normalized Wasserstein Distance (NWD) loss function, leading performance improvements. Specifically, S2-CCFF module enhances information by incorporating an additional layer, while SPDConv downsampling maintains key details reduces computational cost. CSPOK-Fusion mechanism integrates global, local, large branch features, capturing multi-scale representations effectively mitigating interference from backgrounds occlusions, thereby enhancing spatial representation of features across scales. PPA embedded in Backbone network, leverages multi-level feature fusion mechanisms retain strengthen addressing issue loss. NWD focusing on relative positioning shape differences boxes, increases robustness minor perturbations, detection accuracy. Experimental results VisDrone NWPU VHR-10 datasets demonstrate that our approach outperforms state-of-the-art detectors.

Language: Английский

---

Online Traffic Crash Risk Inference Method Using Detection Transformer and Support Vector Machine Optimized by Biomimetic Algorithm

Biomimetics, Journal Year: 2024, Volume and Issue: 9(11), P. 711 - 711

Published: Nov. 19, 2024

Despite the implementation of numerous interventions to enhance urban traffic safety, estimation risk crashes resulting in life-threatening and economic costs remains a significant challenge. In light above, an online inference method for crash based on self-developed TAR-DETR WOA-SA-SVM methods is proposed. The method's robust data capabilities can be applied autonomous mobile robots vehicle systems, enabling real-time road condition prediction, continuous monitoring, timely roadside assistance. First, dataset object detection, named TAR-1, created by extracting information from major roads around Hainan University China incorporating Russian car news. Secondly, we develop innovative Context-Guided Reconstruction Feature Network-based Urban Traffic Objects Detection Model (TAR-DETR). model demonstrates detection accuracy 76.8% objects, which exceeds performance other state-of-the-art models. employed TAR-1 extract features, feature was designated as TAR-2. TAR-2 comprises six features three categories. A new algorithm proposed optimize parameters (C, g) SVM, thereby enhancing robustness inference. developed combining Whale Optimization Algorithm (WOA) Simulated Annealing (SA), Hybrid Bionic Intelligent Algorithm. inputted into Support Vector Machine (SVM) optimized using hybrid used infer crashes. achieves average 80%

Language: Английский

---