Starting from the structure: A review of small object detection based on deep learning

Image and Vision Computing, Год журнала: 2024, Номер 146, С. 105054 - 105054

Опубликована: Апрель 29, 2024

Язык: Английский

---

Drones in Action: A Comprehensive Analysis of Drone-Based Monitoring Technologies

Data & Metadata, Год журнала: 2024, Номер 3

Опубликована: Сен. 2, 2024

Unmanned aerial vehicles (UAVs), commonly referred to as drones, are extensively employed in various real-time applications, including remote sensing, disaster management and recovery, logistics, military operations, search rescue, law enforcement, crowd monitoring control, owing their affordability, rapid processing capabilities, high-resolution imagery. Additionally, drones mitigate risks associated with terrorism, disease spread, temperature fluctuations, crop pests, criminal activities. Consequently, this paper thoroughly analyzes UAV-based surveillance systems, exploring the opportunities, challenges, techniques, future trends of drone technology. It covers common image preprocessing methods for highlights notable one- two-stage deep learning algorithms used object detection drone-captured images. The also offers a valuable compilation online datasets containing drone-acquired photographs researchers. Furthermore, it compares recent imaging detailing purposes, descriptions, findings, limitations. Lastly, addresses potential research directions challenges related usage

Язык: Английский

---

Lightweight Detection Methods for Insulator Self-Explosion Defects

Sensors, Год журнала: 2024, Номер 24(1), С. 290 - 290

Опубликована: Янв. 3, 2024

The accurate and efficient detection of defective insulators is an essential prerequisite for ensuring the safety power grid in new generation intelligent electrical system inspections. Currently, traditional object algorithms detecting images face issues such as excessive parameter size, low accuracy, slow speed. To address aforementioned issues, this article proposes insulator defect model based on lightweight Faster R-CNN (Faster Region-based Convolutional Network) R-CNN-tiny). First, model’s backbone network turned into a version it by substituting EfficientNet ResNet (Residual Network), greatly decreasing parameters while increasing its accuracy. second step to employ feature pyramid build maps with various resolutions fusion, which enables objects at scales. In addition, replacing ordinary convolutions more depth-wise separable increases speed slightly reducing Transfer learning introduced, training method involving freezing unfreezing employed enhance network’s ability detect small target defects. proposed validated using self-exploding dataset. experimental results show that R-CNN-tiny significantly outperforms (ResNet) terms mean average precision (mAP), frames per (FPS), number parameters.

Язык: Английский

---

The Use of Head-Mounted Display Systems for Upper Limb Kinematic Analysis in Post-Stroke Patients: A Perspective Review on Benefits, Challenges and Other Solutions

Bioengineering, Год журнала: 2024, Номер 11(6), С. 538 - 538

Опубликована: Май 24, 2024

In recent years, there has been a notable increase in the clinical adoption of instrumental upper limb kinematic assessment. This trend aligns with rising prevalence cerebrovascular impairments, one most prevalent neurological disorders. Indeed, is growing need for more objective outcomes to facilitate tailored rehabilitation interventions following stroke. Emerging technologies, like head-mounted virtual reality (HMD-VR) platforms, have responded this demand by integrating diverse tracking methodologies. Specifically, HMD-VR technology enables comprehensive body posture, encompassing hand position and gesture, facilitated either through specific tracker placements or via integrated cameras coupled sophisticated computer graphics algorithms embedded within helmet. review aims present state-of-the-art applications platforms analysis post-stroke patients, comparing them conventional systems. Additionally, we address potential benefits challenges associated these platforms. These systems might represent promising avenue safe, cost-effective, portable motor assessment field neurorehabilitation, although other systems, including robots, should be taken into consideration.

Язык: Английский

---

Research on Infrared Dim Target Detection Based on Improved YOLOv8

Remote Sensing, Год журнала: 2024, Номер 16(16), С. 2878 - 2878

Опубликована: Авг. 7, 2024

Addressing the formidable challenges in spatial infrared dim target detection, this paper introduces an advanced detection approach based on refinement of YOLOv8 algorithm. In contrast to conventional YOLOv8, our method achieves remarkable improvements accuracy through several novel strategies. Notably, by incorporating a deformable convolutional module into backbone network, effectively captures more intricate image features, laying solid foundation for subsequent feature fusion and head predictions. Furthermore, dedicated small layer, built upon original model, significantly enhances model’s capability recognizing targets, thereby boosting overall performance. Additionally, we utilize WIoU-v3 as localization regression loss function, reducing sensitivity positional errors leveraging advantages multi-attention mechanisms. To enrich quantity quality dataset, employ enhancement techniques augment dataset. Extensive experiments demonstrate exceptional performance method. Specifically, precision 95.6%, recall rate 94.7%, mean average (mAP) exceeding 97.4%, representing substantial over traditional Moreover, speed reaches 59 frames/s, satisfying requirements real-time detection. This achievement not only validates efficacy superiority algorithm but also offers insights methodologies research applications related fields, holding immense potential future applications.

Язык: Английский

---

SDG-YOLOv8: Single-domain generalized object detection based on domain diversity in traffic road scenes

Displays, Год журнала: 2025, Номер 87, С. 102948 - 102948

Опубликована: Янв. 5, 2025

Язык: Английский

---

Integrating EnlightenGAN for enhancing car logo detection under challenging lighting conditions

Multimedia Tools and Applications, Год журнала: 2025, Номер unknown

Опубликована: Янв. 22, 2025

Язык: Английский

---

An artificial intelligence image-based approach for colloid detection in saturated porous media

Colloids and Surfaces A Physicochemical and Engineering Aspects, Год журнала: 2025, Номер unknown, С. 136503 - 136503

Опубликована: Фев. 1, 2025

Язык: Английский

---

Lightweight obstacle detection for unmanned mining trucks in open-pit mines

Scientific Reports, Год журнала: 2025, Номер 15(1)

Опубликована: Март 16, 2025

This paper aims to solve the problem of difficulty in balancing model size and detection accuracy unmanned mining truck network open-pit mines, as well that existing is not suitable for equipment. To address this problem, we proposed a lightweight vehicle algorithm based on improvement YOLOv8. Through series innovative structural adjustments optimization strategies, has achieved high low complexity. replaces backbone YOLOv8s with FasterNet_t0 (FN) network. advantages simple structure lightweight, which effectively reduces amount calculation parameters model. Then feature extraction YOLOv8 neck replaced BiFPN (Bi-directional Feature Pyramid Network) structure. By increasing cross-layer connections removing nodes contribution fusion, fusion utilization features different scales are optimized, performance further improved, number calculations reduced. make up possible loss caused by improvements, head Dynamic Head. design can introduce self-attention mechanism from three dimensions scale, space, task, significantly improving while avoiding additional computational burden. In terms function, introduces combination SIoU NWD (normalized Gaussian Wasserstein distance) loss. These two enable cope scenarios more accurately, especially effect small target trucks improved. addition, also adopts amplitude-based layer adaptive sparse pruning (LAMP) compress maintaining efficient performance. strategy, its dependence computing resources key experimental part, dataset 3000 images was first constructed, these were preprocessed, including image enhancement, denoising, cropping, scaling. The environment set Autodl cloud server, using PyTorch 2.5.1 framework Python 3.10 environment. four sets ablation experiments, verified specific impact each results show strategy improves model, greatly reducing Finally, conducted comprehensive comparative analysis improved other popular algorithms models. our leads 76.9%, than 10% higher similar At same time, compared models achieve levels, only about 20% size. fully prove adopted feasible obvious efficiency.

Язык: Английский

---

Wasp‐Hive Candidate Site Search System Using a Small Drone

Entomological Research, Год журнала: 2025, Номер 55(3)

Опубликована: Март 1, 2025

ABSTRACT Early detection of wasp hives is crucial for mitigating their impact on native species, preventing agricultural damage, and improving pest control strategies. Traditional methods rely ground surveys sensor‐based tracking individual insects, which are often labor‐intensive, time‐consuming, prone to errors because environmental constraints. The integration artificial intelligence drone‐based imaging has the potential revolutionize ecological monitoring by providing scalable, efficient, noninvasive detecting hives. However, research AI‐assisted hive remains limited, with most studies focusing large‐scale wildlife rather than small‐object localization. Therefore, we propose a system searching candidate site using small drone. In proposed system, drone equipped camera takes aerial images error range. Subsequently, three‐dimensional (3D) modeling performed captured 3D surveying toolkit, deep learning–based completed model extract GPS information detected target.

Язык: Английский

---