Small Object Detection Research Articles

ATBHC-YOLO: aggregate transformer and bidirectional hybrid convolution for small object detection

ATBHC-YOLO: aggregate transformer and bidirectional hybrid convolution for small object detection

GreenFruitDetector: Lightweight green fruit detector in orchard environment.

Detecting green fruits presents significant challenges due to their close resemblance in color to the leaves in an orchard environment. We designed GreenFruitDetector, a lightweight model based on an improved YOLO v8 architecture, specifically for green fruit detection. In the Backbone network, we replace ordinary convolution with Deformable Convolution to enhance the extraction of geometric features. Additionally, we designed MCAG-DC (Multi-path Coordinate Attention Guided Deformer Convolution) to replace the convolution in C2f, enhancing the Backbone's feature extraction capability when encountering occlusion problems. For the Neck part of the algorithm, we designed a Fusion-neck structure that integrates spatial detail information from feature maps at different scales, thereby enhancing the network's ability to extract multi-scale information. Additionally, we devised a new detection head that incorporates multi-scale information, significantly improving the detection of small and distant objects. Finally, we applied channel pruning techniques to reduce the model size, parameter count, and FLOPs to 50%, 55%, and 44% of the original, respectively. We trained and evaluated the improved model on three green fruit datasets. The accuracy of the improved model reached 94.5%, 84.4%, and 85.9% on the Korla Pear, Guava, and Green Apple datasets, respectively, representing improvements of 1.17%, 1.1%, and 1.77% over the baseline model. The mAP@0.5 increased by 0.72%, 6.5%, and 0.9%, respectively, and the recall rate increased by 1.97%, 1.1%, and 0.49%, respectively.

Mamba-UAV-SegNet: A Multi-Scale Adaptive Feature Fusion Network for Real-Time Semantic Segmentation of UAV Aerial Imagery

Accurate semantic segmentation of high-resolution images captured by unmanned aerial vehicles (UAVs) is crucial for applications in environmental monitoring, urban planning, and precision agriculture. However, challenges such as class imbalance, small-object detection, and intricate boundary details complicate the analysis of UAV imagery. To address these issues, we propose Mamba-UAV-SegNet, a novel real-time semantic segmentation network specifically designed for UAV images. The network integrates a Multi-Head Mamba Block (MH-Mamba Block) for enhanced multi-scale feature representation, an Adaptive Boundary Enhancement Fusion Module (ABEFM) for improved boundary-aware feature fusion, and an edge-detail auxiliary training branch to capture fine-grained details. The practical utility of our method is demonstrated through its application to farmland segmentation. Extensive experiments on the UAV-City, VDD, and UAVid datasets show that our model outperforms state-of-the-art methods, achieving mean Intersection over Union (mIoU) scores of 71.2%, 77.5%, and 69.3%, respectively. Ablation studies confirm the effectiveness of each component and their combined contributions to overall performance. The proposed method balances segmentation accuracy and computational efficiency, maintaining real-time inference speeds suitable for practical UAV applications.

A Reparameterization Feature Redundancy Extract Network for Unmanned Aerial Vehicles Detection

In unmanned aerial vehicles (UAVs) detection, challenges such as occlusion, complex backgrounds, motion blur, and inference time often lead to false detections and missed detections. General object detection frameworks encounter difficulties in adequately tackling these challenges, leading to substantial information loss during network downsampling, inadequate feature fusion, and being unable to meet real-time requirements. In this paper, we propose a Real-Time Small Object Detection YOLO (RTSOD-YOLO) model to tackle the various challenges faced in UAVs detection. We further enhance the adaptive nature of the Adown module by incorporating an adaptive spatial attention mechanism. This mechanism processes the downsampled feature maps, enabling the model to better focus on key regions. Secondly, to address the issue of insufficient feature fusion, we employ combined serial and parallel triple feature encoding (TFE). This approach fuses scale-sequence features from both shallow features and twice-encoded features, resulting in a new small-scale object detection layer. While enhancing the global context awareness of the existing detection layers, this also enriches the small-scale object detection layer with detailed information. Since rich redundant features often ensure a comprehensive understanding of the input, which is a key characteristic of deep neural networks, we propose a more efficient redundant feature generation module. This module generates more feature maps with fewer parameters. Additionally, we introduce reparameterization techniques to compensate for potential feature loss while further improving the model’s inference speed. Experimental results demonstrate that our proposed RTSOD-YOLO achieves superior detection performance, with mAP50/mAP50:95 reaching 97.3%/51.7%, which represents improvement of 3%/3.5% over YOLOv8, and 2.6%/0.1% higher than YOLOv10. Additionally, it has the lowest parameter count and FLOPs, making it highly efficient in terms of computational resources.

Object Detection and Tracking in Maritime Environments in Case of Person-Overboard Scenarios: An Overview

We examine the current state of the art and the related research on the automated detection and tracking of small objects—or persons—in the context of a person-overboard (POB) scenario and present the associated governing relationship between different technologies, platforms, and approaches as a system of systems. A novel phase model, structuring a POB scenario, comprises three phases: (1) detection, (2) search and track, and (3) rescue. Within these phases, we identify the central areas of responsibility and describe in detail the phases (1) and (2). We emphasize the importance of a high-level representation of different systems and their interactions to comprehensively represent the complexity and dynamics of POB scenarios. Our systematic classification and detailed description of the technologies and methods used provide valuable insights to support future regulatory and research activities. Our primary aim is to advance the development of corresponding technologies and standards.

CDANet: a small object detection model for unmanned aerial vehicle images based on cross-layer guidance and skip dilated fusion

CDANet: a small object detection model for unmanned aerial vehicle images based on cross-layer guidance and skip dilated fusion

Vision-guided crack identification and size quantification framework for dam underwater concrete structures

Remotely operated vehicles with cameras provide a non-contact inspection solution for dam underwater structural defect detection. However, manual information extraction methods suffer from problems like labor costs and high misjudgment. This study proposes an integrated dam underwater crack identification and size quantification framework using machine vision and deep learning. First, a real-time automatic detection framework for dam underwater cracks is built via You Only Look Once v5 (YOLOv5), and four different backbone detectors are introduced to balance detection accuracy and speed. Then, the Swin-Transformer module is inserted into the YOLOv5 model to enhance its feature extraction capability and small object detection capability. Next, a method for measuring the true size of cracks was constructed based on deep learning and infrared laser rangefinders. In this study, physical model experiments and actual engineering projects are combined to validate the generalization capability of the proposed crack identification and size quantification method. Experimental results show that the Swin-Transformer-based YOLOv5 model effectively balances detection accuracy and speed with a precision of 0.986, a recall of 0.979, a mean average precision of 0.985, and a frame rate of 68 frames per second in detecting underwater crack images with 768 × 576 pixels. In addition, the proposed method can accurately identify and detect cracks in complex underwater scenes, including obstacle interference, tilt shooting angle, uneven illumination, and turbid water scenarios. Moreover, the method proposed in this paper can quantify the overall size and geometric parameters of underwater cracks with relatively small errors.

Efficient Small Object Detection You Only Look Once: A Small Object Detection Algorithm for Aerial Images.

Aerial images have distinct characteristics, such as varying target scales, complex backgrounds, severe occlusion, small targets, and dense distribution. As a result, object detection in aerial images faces challenges like difficulty in extracting small target information and poor integration of spatial and semantic data. Moreover, existing object detection algorithms have a large number of parameters, posing a challenge for deployment on drones with limited hardware resources. We propose an efficient small-object YOLO detection model (ESOD-YOLO) based on YOLOv8n for Unmanned Aerial Vehicle (UAV) object detection. Firstly, we propose that the Reparameterized Multi-scale Inverted Blocks (RepNIBMS) module is implemented to replace the C2f module of the Yolov8n backbone extraction network to enhance the information extraction capability of small objects. Secondly, a cross-level multi-scale feature fusion structure, wave feature pyramid network (WFPN), is designed to enhance the model's capacity to integrate spatial and semantic information. Meanwhile, a small-object detection head is incorporated to augment the model's ability to identify small objects. Finally, a tri-focal loss function is proposed to address the issue of imbalanced samples in aerial images in a straightforward and effective manner. In the VisDrone2019 test set, when the input size is uniformly 640 × 640 pixels, the parameters of ESOD-YOLO are 4.46 M, and the average mean accuracy of detection reaches 29.3%, which is 3.6% higher than the baseline method YOLOv8n. Compared with other detection methods, it also achieves higher detection accuracy with lower parameters.

Dynamic feature and context enhancement network for faster detection of small objects

While traditional object detection methods have achieved significant success in recent years, their performance in detecting small objects in aerial remote sensing images remains unsatisfactory. Small objects often occupy only a few pixels, leading to a loss of fine-grained information. This paper proposes a dynamic feature and context enhancement network (DFCE) to address noise and pixel-level region weighting issues in small object detection. The DFCE network effectively detects small objects by dynamically selecting features within regions and establishing connections between local and global contextual information. The introduced dynamic multi-dimensional attention (DMA) module selects key information via a crossover mechanism and assigns different weights to highlight important features. Based on DMA, the regional feature processing (RFP) module and multi-dimensional pool transformer (MPT) module are developed to capture key information and contextual information, respectively. Experimental results demonstrate that the DFCE network improves average precision (AP) by 3.1%, 9%, and 2.2% on two remote sensing datasets and one conventional dataset, achieving an inference speed of 30 frames per second (FPS). Given these advancements, the DFCE model’s powerful key feature extraction and contextual association capabilities show strong potential for broader applications, including sign language recognition, defect detection in industrial settings, and more.

DeployFusion: A Deployable Monocular 3D Object Detection with Multi-Sensor Information Fusion in BEV for Edge Devices.

To address the challenges of suboptimal remote detection and significant computational burden in existing multi-sensor information fusion 3D object detection methods, a novel approach based on Bird's-Eye View (BEV) is proposed. This method utilizes an enhanced lightweight EdgeNeXt feature extraction network, incorporating residual branches to address network degradation caused by the excessive depth of STDA encoding blocks. Meantime, deformable convolution is used to expand the receptive field and reduce computational complexity. The feature fusion module constructs a two-stage fusion network to optimize the fusion and alignment of multi-sensor features. This network aligns image features to supplement environmental information with point cloud features, thereby obtaining the final BEV features. Additionally, a Transformer decoder that emphasizes global spatial cues is employed to process the BEV feature sequence, enabling precise detection of distant small objects. Experimental results demonstrate that this method surpasses the baseline network, with improvements of 4.5% in the NuScenes detection score and 5.5% in average precision for detection objects. Finally, the model is converted and accelerated using TensorRT tools for deployment on mobile devices, achieving an inference time of 138 ms per frame on the Jetson Orin NX embedded platform, thus enabling real-time 3D object detection.

A Heatmap-Supplemented R-CNN Trained Using an Inflated IoU for Small Object Detection

Object detection architectures struggle to detect small objects across applications including remote sensing and autonomous vehicles. Specifically, for unmanned aerial vehicles, poor detection of small objects directly limits this technology’s applicability. Objects both appear smaller than they are in large-scale images captured in aerial imagery and are represented by reduced information in high-altitude imagery. This paper presents a new architecture, CR-CNN, which predicts independent regions of interest from two unique prediction branches within the first stage of the network: a conventional R-CNN convolutional backbone and an hourglass backbone. Utilizing two independent sources within the first stage, our approach leads to an increase in successful predictions of regions that contain smaller objects. Anchor-based methods such as R-CNNs also utilize less than half the number of small objects compared to larger ones during training due to the poor intersection over union (IoU) scores between the generated anchors and the groundtruth—further reducing their performance on small objects. Therefore, we also propose artificially inflating the IoU of smaller objects during training using a simple, size-based Gaussian multiplier—leading to an increase in the quantity of small objects seen per training cycle based on an increase in the number of anchor–object pairs during training. This architecture and training strategy led to improved detection overall on two challenging aerial-based datasets heavily composed of small objects while predicting fewer false positives compared to Mask R-CNN. These results suggest that while new and unique architectures will continue to play a part in advancing the field of object detection, the training methodologies and strategies used will also play a valuable role.

MS-YOLO: A Lightweight and High-Precision YOLO Model for Drowning Detection.

A novel detection model, MS-YOLO, is developed in this paper to improve the efficiency of drowning rescue operations. The model is lightweight, high in precision, and applicable for intelligent hardware platforms. Firstly, the MD-C2F structure is built to capture the subtle movements and posture changes in various aquatic environments, with a light weight achieved by introducing dynamic convolution (DcConv). To make the model perform better in small object detection, the EMA mechanism is incorporated into the MD-C2F. Secondly, the MSI-SPPF module is constructed to improve the performance in identifying the features of different scales and the understanding of complex backgrounds. Finally, the ConCat single-channel fusion is replaced by BiFPN weighted channel fusion to retain more feature information and remove the irrelevant information in drowning features. Relative to the Faster R-CNN, SSD, YOLOv6, YOLOv9, and YOLOv10, the MS-YOLO achieves an average accuracy of 86.4% in detection on a self-built dataset at an ultra-low computational cost of 7.3 GFLOPs.

The Detection and Classification of Scaphoid Fractures in Radiograph by Using a Convolutional Neural Network.

Scaphoid fractures, particularly occult and non-displaced fractures, are difficult to detect using traditional X-ray methods because of their subtle appearance and variability in bone density. This study proposes a two-stage CNN approach to detect and classify scaphoid fractures using anterior-posterior (AP) and lateral (LA) X-ray views for more accurate diagnosis. This study emphasizes the use of multi-view X-ray images (AP and LA views) to improve fracture detection and classification. The multi-view fusion module helps integrate information from both views to enhance detection accuracy, particularly for occult fractures that may not be visible in a single view. The proposed method includes two stages, which are stage 1: detect the scaphoid bone using Faster RCNN and a Feature Pyramid Network (FPN) for region proposal and small object detection. The detection accuracy for scaphoid localization is 100%, with Intersection over Union (IoU) scores of 0.8662 for AP views and 0.8478 for LA views. And stage 2: perform fracture classification using a ResNet backbone and FPN combined with a multi-view fusion module to combine features from both AP and LA views. This stage achieves a classification accuracy of 89.94%, recall of 87.33%, and precision of 90.36%. The proposed model performs well in both scaphoid bone detection and fracture classification. The multi-view fusion approach significantly improves recall and accuracy in detecting fractures compared to single-view approaches. In scaphoid detection, both AP and LA views achieved 100% detection accuracy. In fracture detection, using multi-view fusion, the accuracy for AP views reached 87.16%, and for LA views, it reached 83.83%. The multi-view fusion model effectively improves the detection of scaphoid fractures, particularly in cases of occult and non-displaced fractures. The model provides a reliable, automated approach to assist clinicians in detecting and diagnosing scaphoid fractures more efficiently.

YOLO-DHGC: Small Object Detection Using Two-Stream Structure with Dense Connections.

Small object detection, which is frequently applied in defect detection, medical imaging, and security surveillance, often suffers from low accuracy due to limited feature information and blurred details. This paper proposes a small object detection method named YOLO-DHGC, which employs a two-stream structure with dense connections. Firstly, a novel backbone network, DenseHRNet, is introduced. It innovatively combines a dense connection mechanism with high-resolution feature map branches, effectively enhancing feature reuse and cross-layer fusion, thereby obtaining high-level semantic information from the image. Secondly, a two-stream structure based on an edge-gated branch is designed. It uses higher-level information from the regular detection stream to eliminate irrelevant interference remaining in the early processing stages of the edge-gated stream, allowing it to focus on processing information related to shape boundaries and accurately capture the morphological features of small objects. To assess the effectiveness of the proposed YOLO-DHGC method, we conducted experiments on several public datasets and a self-constructed dataset. Exceptionally, a defect detection accuracy of 96.3% was achieved on the Market-PCB public dataset, demonstrating the effectiveness of our method in detecting small object defects for industrial applications.

BGF-YOLOv10: Small Object Detection Algorithm from Unmanned Aerial Vehicle Perspective Based on Improved YOLOv10.

With the rapid development of deep learning, unmanned aerial vehicles (UAVs) have acquired intelligent perception capabilities, demonstrating efficient data collection across various fields. In UAV perspective scenarios, captured images often contain small and unevenly distributed objects, and are typically high-resolution. This makes object detection in UAV imagery more challenging compared to conventional detection tasks. To address this issue, we propose a lightweight object detection algorithm, BGF-YOLOv10, specifically designed for small object detection, based on an improved version of YOLOv10n. First, we introduce a novel YOLOv10 architecture tailored for small objects, incorporating BoTNet, variants of C2f and C3 in the backbone, along with an additional small object detection head, to enhance detection performance for small objects. Second, we embed GhostConv into both the backbone and head, effectively reducing the number of parameters by nearly half. Finally, we insert a Patch Expanding Layer module in the neck to restore the feature spatial resolution. Experimental results on the VisDrone-DET2019 and UAVDT datasets demonstrate that our method significantly improves detection accuracy compared to YOLO series networks. Moreover, when compared to other state-of-the-art networks, our approach achieves a substantial reduction in the number of parameters.

SOD-YOLO: A lightweight small object detection framework

Currently, lightweight small object detection algorithms for unmanned aerial vehicles (UAVs) often employ group convolutions, resulting in high Memory Access Cost (MAC) and rendering them unsuitable for edge devices that rely on parallel computing. To address this issue, we propose the SOD-YOLO model based on YOLOv7, which incorporates a DSDM-LFIM backbone network and includes a small object detection branch. The DSDM-LFIM backbone network, which combines Deep-Shallow Downsampling Modules (DSD Modules) and Lightweight Feature Integration Modules (LFI Modules), avoids excessive use of group convolutions and element-wise operations. The DSD Module focuses on extracting both deep and shallow features from feature maps using fewer parameters to obtain richer feature representations. The LFI Module, is a dual-branch feature integration module designed to consolidate feature information. Experimental results demonstrate that the SOD-YOLO model achieves an AP50 of 50.7% and a FPS of 72.5 on the VisDrone validation set. Compared to YOLOv7, our model reduces computational costs by 20.25% and decreases the number of parameters by 17.89%. After scaling the number of channels in the model, it achieves an AP50 of 33.4% with an inference time of 27.3ms on the Atlas 200I DK A2. These experimental results indicate that the SOD-YOLO model can effectively perform small object detection tasks in a large number of aerial images captured by UAVs.

A Dual Detection Head YOLO Model With Its Application in Wheat Ear Recognition

The spike number(SN) per unit area is an important indicator for measuring wheat yield. The use of AI to detect and count wheat ears is increasingly valued by researchers. In order to improve the accuracy of wheat ear recognition, the model is deployed on limited mobile devices to achieve accurate estimation of wheat yield. This paper proposes a lightweight hybrid design network Wheat-YoloNet based on improved YOLOv8 for wheat ear recognition. On the basis of the YOLOv8 network structure, we only retain small and large object detection heads, reducing the number of network parameters while ensuring detection performance. We added CBAM attention mechanism module to Backbone to improve the feature extraction performance of the model and introduce Wise-IoU to balance the samples. Compared with the YOLOv8n baseline model, the improved model has reduced parameter count by 33.3%, outperforming other comparison object detection algorithms, and can be deployed and run on devices with limited computing power.

AI-empowered Search Drone: Advanced AI-based Tiny Object Search Drone for Broad Areas

Abstract. Detecting the garbage distribution and position in broad terrains(beaches, grasslands, squares, etc) is extremely important for environmental protection and urban image. Manually screen these terrains is time-consuming and ineicient. Autonomous robot may find the wastes automatically, but they may hinder and even endanger the pedestrians. Unmanned aerial systems(UAS) are ideal tools for monitoring broad areas and detecting garbage. In this project we create an AI-empowered Search Drone for extensive garbage search in broad regions. The camera on our drone and the external image transmission module on the computer receive and process top-view image from the sky. The images from high altitude lead to the reduction in target size and object distortion. We innovately modified the network structure of the oicial YOLOv5 model by adding layers specifically for small object detection, enhancing its adaptability to small targets captured by drones.To address camera image distortion, we leveraged the OpenCV (cv2) library to obtain and apply a distortion correction matrix. Furthermore, we devised an algorithm that calculates the real-world location of targets based on their pixel positions, the drones height, and the ratio between pixel and real-world dimensions.To strengthen the models garbage search capabilities, we simulated various garbage search scenarios in open environments and collected image datasets using drones for specialized training. Finally, we utilized the PyQt library to develop a simplistic front-end interface for data analysis and comparison of the detection performance between the original and optimized models. Experimental results demonstrate that the optimized model outperforms the original model, effectively enhancing garbage search eiciency in extensive areas. This project represents a significant step forward in leveraging technology to combat environmental pollution.

CSSDet: small object detection via cross-scale feature enhancement on drone-view images

ABSTRACT Object detection on drone-view images is vital for applications like intelligent transportation, abnormal behavior detection, and urban surveillance. However, the diverse perspectives and altitudes from which Unmanned Aerial Vehicle (UAV) capture scenes pose challenges due to multi-scale differences and small object sizes. To address this, we propose CSSDet, emphasizing multi-scale object detection and small object enhancement. CSSDet integrates an additional detection head for small objects and employs a bidirectional weighted feature pyramid network for effective cross-scale feature fusion, enhancing global perceptual capability. Additionally, we introduce coordinate attention and Involution modules to mitigate information loss. Experiments are conducted on two publicly available datasets, VisDrone and UAVDT. Quantitative results from the experiments indicate that the proposed CSSDet achieves optimal performance across a wide range of evaluation metrics, with particular strengths demonstrated in multi-scale object detection and small-size object detection. Qualitative results showcase its remarkable effectiveness in various aspects of object detection, including small object detection, objects at different scales and categories, multi-perspective scenarios, and complex scenes. The substantial experimental findings confirm that CSSDet attains comprehensive object detection performance on drone-view images. The source code is available at https://github.com/Gui-Cheng/CSSDet.

YOLO-ESL: An Enhanced Pedestrian Recognition Network Based on YOLO

Pedestrian detection is a critical task in computer vision; however, mainstream algorithms often struggle to achieve high detection accuracy in complex scenarios, particularly due to target occlusion and the presence of small objects. This paper introduces a novel pedestrian detection algorithm, YOLO-ESL, based on the YOLOv7 framework. YOLO-ESL integrates the ELAN-SA module, designed to enhance feature extraction, with the LGA module, which improves feature fusion. The ELAN-SA module optimizes the flexibility and efficiency of small object feature extraction, while the LGA module effectively integrates multi-scale features through local and global attention mechanisms. Additionally, the CIOUNMS algorithm addresses the issue of target loss in cases of high overlap, improving boundary box filtering. Evaluated on the VOC2012 pedestrian dataset, YOLO-ESL achieved an accuracy of 93.7%, surpassing the baseline model by 3.0%. Compared to existing methods, this model not only demonstrates strong performance in handling occluded and small object detection but also remarkable robustness and efficiency.