YOLO‐Tiny: A lightweight small object detection algorithm for UAV aerial imagery

Monitoring the critically endangered Amur tiger (Panthera tigris altaica) is crucial for the conservation of the natural environment. This paper proposes a novel lightweight deep‐learning network for detecting Amur tigers in mobile and resource‐constrained environments, like those using unmanned aerial vehicles. The network uses an anchor‐free mechanism with modified CSPNet and cross stage partial (CSP)‐path aggregation network (PAN) structures, which improves the model's feature extraction capabilities. Label assignment strategy is also improved to stabilize the model training process. Additionally, the random Mosaic and Mixup data augmentation strategy is utilized to address the overfitting issue due to insufficient data in the data set. The model achieves 55.5% mean average precision (mAP [0.5:0.95]) with only 0.617 million parameters and 73.58 frames per second on mobile CPU with pixels input. Results show that the model is an accurate, fast, and practical detector of the Amur tiger, serving as a reference for wildlife detectors of other species.

Apple detection plays a critical role in enabling the functionality of harvesting robots within natural orchard environments. To address challenges related to low detection accuracy, slow inference speed, and high parameter count, we present PcMNet, a lightweight detection model based on an improved YOLOv8 network. Initially, we employed Partial Convolution (Pconv) to construct a PR module, forming the Pconv-block, which subsequently replaced the original C2f feature extraction module within the YOLOv8n backbone. This replacement led to improvements in both detection accuracy and speed, while simultaneously reducing computational complexity (FLOPs), parameter count, and model size. Furthermore, the Cross-Scale Feature Fusion (CCFF) module was refined into Faster-Cross-Scale Feature Fusion (Faster-CCFF) with the integration of Pconv-block, significantly enhancing the model's feature extraction and fusion capabilities. Additionally, we introduced Mixed Local Channel Attention (MLCA) to further strengthen the model's capacity to capture essential features while effectively suppressing background noise. Experimental results demonstrate that PcMNet achieved a detection accuracy of 92.8 % and an mAP@0.5 of 95.5 %, representing improvements of 1.4 and 0.7 percentage points, respectively, over YOLOv8n. Moreover, PcMNet successfully reduced FLOPs, parameter count, and model size to 5.1 G, 1.4 M, and 3.2 MB, respectively. The per-image detection time was reduced to 2.3 ms, indicating reductions of 37.80 %, 53.33 %, 49.21 %, and 56.60 % in FLOPs, parameters, model size, and detection time compared to YOLOv8n. When deployed on edge computing devices with TensorRT acceleration, PcMNet achieved a detection rate of 92 FPS. Field validation experiments conducted in natural orchard environments confirmed PcMNet's superior ability to detect apples under challenging conditions, such as occlusions and varying lighting conditions. Its lightweight design and rapid detection capabilities provide a valuable reference for achieving real-time apple detection in automated and intelligent harvesting robots, thereby contributing to advancements in smart agriculture.

Object detection on unmanned aerial vehicle (UAV) images has been a hot research topic recently. However, object detection models for general scenarios struggle with UAV images due to the challenges of detecting small targets and handling complex image backgrounds. To solve the two issues, we proposed DPH-YOLOv8, an enhanced version of YOLOv8 tailored for UAV scenarios. First, we improved the prediction head from three prediction heads to double prediction heads (DPHs), reducing the model’s parameters by 42.9% while improving the mAP for small targets. Second, we designed a tiny-path feature fusion (TP-Fusion) module to fuse richer detailed information, enabling the detector to accurately match targets of different sizes and shapes. Third, we introduced a coordinate attention (CA) module to help the model reduce the interference of complex background information and focus on detecting foreground targets. Finally, bottleneck modules were added before the prediction heads to enhance the extraction of small target features. Extensive experimental results on both the VisDrone2021 and UAVDT benchmarks demonstrated that DPH-YOLOv8 not only improved the mAP by 4.5% on VisDrone2021 and 2.2% on UAVDT but also reduced localization error by 0.53%, confusion with objects by 0.69%, and confusion with background by 0.45%. These enhancements, along with a reduction in the model’s parameters, make DPH-YOLOv8 more suitable for UAV scenarios.

Unmanned Aerial Vehicle (UAV) image object detection plays a crucial role in various fields. However, compared with natural images, UAV images are characterized by significant target scale variations, complex backgrounds, dense small targets, and clustered target distributions, which pose serious challenges to object detection tasks. To address these issues, this study proposes an improved object detection algorithm named MFRA-YOLO, which combines multi-scale feature fusion and receptive-field attention-based convolution, built upon the baseline YOLOv8n algorithm. First, Monte Carlo attention is integrated into receptive-field attention-based convolution to enhance cross-scale information interaction capability, thereby forming a novel convolutional module for downsampling operations. Second, a multi-scale selective fusion module is incorporated into the feature fusion network to enable adaptive cross-scale integration of features. When coupled with the scale sequence feature fusion module, this integration significantly enhances the detection performance for small targets. Finally, the Focaler-PIoUv2 loss function is designed to replace the CIoU in the baseline algorithm. This replacement allows the algorithm to better balance hard and easy samples and improve detection accuracy. Experimental results on the public dataset VisDrone2019 show that MFRA-YOLO achieves superior accuracy-efficiency tradeoffs compared to the baseline YOLOv8n and other YOLO variants. Compared to YOLOv8n, MFRA-YOLO improves mAP50 by 3.5% and mAP50:95 by 2.3%, respectively. These performance gains are accompanied by only modest increases in parameter count and computational cost. Notably, while maintaining performance at 143 FPS, thereby satisfying the real-time deployment requirements for UAVs. Furthermore, compared with several state-of-the-art algorithms designed for UAV scenarios, MFRA-YOLO also offers distinct advantages. To verify the generalization ability and stability of MFRA-YOLO, comparative experiments are carried out on the RSOD dataset, where our algorithm consistently demonstrates excellent detection performance. Overall, these results confirm that MFRA-YOLO not only improves the detection performance for UAV imagery substantially but also achieves an excellent balance between detection accuracy and efficiency.

To address the challenges posed by complex backgrounds and the low occurrence in photovoltaic cell images captured by industrial sensors, we propose a novel defect detection method: MRA-YOLOv8. First, a multi-branch coordinate attention network (MBCANet) is introduced into the backbone. The coordinate attention network (CANet) is incorporated to mitigate the noise impact of background information on the detection task, and multiple branches are employed to enhance the model's feature extraction capability. Second, we integrate a multi-path feature extraction module, ResBlock, into the neck. This module provides finer-grained multi-scale features, improving feature extraction from complex backgrounds and enhancing the model's robustness. Finally, we implement alpha-minimum point distance-based IoU (AMPDIoU) to the head. This loss function enhances the accuracy and robustness of small object detection by integrating minimum point distance-based IoU (MPDIoU) and Alpha-IoU methods. The results demonstrate that MRA-YOLOv8 outperforms other mainstream methods in detection performance. On the photovoltaic electroluminescence anomaly detection (PVEL-AD) dataset, the proposed method achieves a mAP50 of 91.7%, representing an improvement of 3.1% over YOLOv8 and 16.1% over detection transformer (DETR). On the SPDI dataset, our method achieves a mAP50 of 69.3%, showing a 2.1% improvement over YOLOv8 and a 6.6% improvement over DETR. The proposed MRA-YOLOv8 also exhibits great deployment potential. It can be effectively integrated with drone-based inspection systems, allowing for efficient and accurate PV plant inspections. Moreover, to tackle the issue of data imbalance, we propose generating synthetic defect data via generative adversarial networks (GANs), which can supplement the limited defect samples and improve the model's generalization ability.

The detection of small infrared targets lacking texture and shape information in the presence of complex background clutter is a challenge that has attracted considerable research attention in recent years. Typical deep learning-based target detection methods are designed with deeper network structures, which may lose targets in the deeper layers and cannot directly be used for small infrared target detection. Therefore, we designed the attention fusion feature pyramid network (AFFPN) specifically for small infrared target detection. Specifically, it consists of feature extraction and feature fusion modules. In the feature extraction stage, the global contextual prior information of small targets is first considered in the deep layer of the network using the atrous spatial pyramid pooling module. Subsequently, the spatial location and semantic information features of small infrared targets in the shallow and deep layers are adaptively enhanced by the designed attention fusion module to improve the feature representation capability of the network for targets. Finally, high-performance detection is achieved through the multilayer feature fusion mechanism. Moreover, we performed a comprehensive ablation study to evaluate the effectiveness of each component. The results demonstrate that the proposed method performs better than state-of-the-art methods on a publicly available dataset. Furthermore, AFFPN was deployed on an NVIDIA Jetson AGX Xavier development board and achieved real-time target detection, further advancing practical research and applications in the field of unmanned aerial vehicle infrared search and tracking.

Target detection based on unmanned aerial vehicle (UAV) images has increasingly become a hot topic with the rapid development of UAVs and related technologies. UAV aerial images often feature a large number of small targets and complex backgrounds due to the UAV’s flying height and shooting angle of view. These characteristics make the advanced YOLOv4 detection method lack outstanding performance in UAV aerial images. In light of the aforementioned problems, this study adjusted YOLOv4 to the image’s characteristics, making the improved method more suitable for target detection in UAV aerial images. Specifically, according to the characteristics of the activation function, different activation functions were used in the shallow network and the deep network, respectively. The loss for the bounding box regression was computed using the EIOU loss function. Improved Efficient Channel Attention (IECA) modules were added to the backbone. At the neck, the Spatial Pyramid Pooling (SPP) module was replaced with a pyramid pooling module. At the end of the model, Adaptive Spatial Feature Fusion (ASFF) modules were added. In addition, a dataset of forklifts based on UAV aerial imagery was also established. On the PASCAL VOC, VEDAI, and forklift datasets, we ran a series of experiments. The experimental results reveal that the proposed method (YOLO-DRONE, YOLOD) has better detection performance than YOLOv4 for the aforementioned three datasets, with the mean average precision (mAP) being improved by 3.06%, 3.75%, and 1.42%, respectively.

Aiming at the problems of leakage, misdetection and low detection efficiency of transmission line foreign object detection in railroad environment, we propose an improved railroad transmission line foreign object detection algorithm MRS-YOLO based on YOLOv11n. Firstly, a multi-scale Adaptive Kernel Depth Feature Fusion (MAKDF) module is proposed and fused with the C3k2 module to form C3k2_MAKDF, which enhances the model's feature extraction capability for foreign objects of different sizes and shapes. Secondly, a novel Re-calibration Feature Fusion Pyramid Network(RCFPN) is designed as a Neck structure to improve the feature fusion capability of the model. Then, Spatial and Channel Reconstruction Detect Head (SC_Detect) based on spatial and channel preprocessing is designed to improve the overall detection accuracy of the model. Finally, the channel pruning technique is used to reduce the redundancy of the improved model, drastically reduce Parameters and GFLOPs, and improve the detection efficiency. The experimental results show that the mAP50 and mAP50:95 of the MRS-YOLO algorithm proposed in this paper are improved to 94.8% and 86.4%, respectively, which are 0.7 and 2.3 percentage points higher compared to the baseline, while Parameters and GFLOPs are reduced by 44.2% and 17.5%, respectively. It is demonstrated that the improved algorithm can be better applied to the task of foreign object detection in railroad transmission lines.

A power line segmentation model in aerial images based on an efficient multibranch concatenation network

Although the data fusion of hyperspectral images (HSI) and light detection and ranging (LiDAR) has provided significant gains for land-cover classification, it also brings technical obstacles (i.e., it is difficult to capture discriminative local and global spatial-spectral from redundant data and build interactions between heterogeneous data). In this paper, a classification network named enhanced spectral attention and adaptive spatial learning guided network (ESASNet) is proposed for the joint use of HSI and LiDAR. Specifically, first, by combining a convolutional neural network (CNN) with the transformer, adaptive spatial learning (ASL) and enhanced spectral learning (ESL) are proposed to learn the spectral-spatial features from the HSI data and the elevation features from the LiDAR data in the local and global receptive field. Second, considering the characteristics of HSI with a continuous, narrowband spectrum, ESL is designed by adding enhanced local self-attention to enhance the mining of the spectral correlations across the adjacent spectrum. Finally, a feature fusion module is proposed to ensure an efficient information exchange between HSI and LiDAR during spectral features and spatial feature fusion. Experimental evaluations on the HSI-LiDAR dataset clearly illustrate that ESASNet performs better in feature extraction than the state-of-the-art methods. The code is available at https://github.com/AirsterMode/ESASNet.

With the wide application of unmanned aerial vehicles (UAVs), UAV thermal imaging technology has been brought to the attention of numerous experts and scholars. At present, the application of UAV thermal imaging technology is limited by the problem of low resolution of the images captured by the thermal imaging system itself. In order to solve this problem, this paper proposes a UAV thermal image super-resolution reconstruction method based on bimodal image feature fusion, which improves the UAV thermal image super-resolution characterization capability by making use of the feature information provided by the UAV visible image to compensate for the lack of texture and edge information perception in thermal images. The proposed methodology consists of two tasks: The first task involves a bimodal image feature fusion network that extracts texture and edge information from visible images. This addresses the lack of detailed feature information in thermal images with a differential complementary modal feature fusion module, ultimately producing a fused image with high-quality feature information. The second task entails the integration of a gated multilevel feature fusion module into the UAV thermal image reconstruction network as an attention mechanism, thereby improving the use of fused image feature information for generating high-resolution UAV thermal images. Finally, a joint adaptive training strategy is utilized to further enhance the super-resolution effect of UAV thermal images. The experimental results show that the proposed method can effectively extract the visible image information to compensate for the characterization information in the UAV thermal image super-resolution, and achieves the optimal values of 31.53 dB and 0.9116 in PSNR and SSIM metrics, respectively, providing a new solution for the thermal image super-resolution of UAVs.

As an important cultural heritage of China and a world treasure, the Dunhuang murals have undergone thousands of years of weathering, oxidation, and human interference, resulting in extensive damage such as cracks, peeling, and fading on their surfaces. Additionally, due to their complex colour palette, it is challenging to quickly identify damaged areas, leading to a situation where the rate of damage to the Dunhuang murals far exceeds the rate of restoration. To achieve efficient and precise monitoring of damage to Dunhuang murals and support cultural heritage preservation efforts, this paper proposes an improved lightweight, high-precision method for monitoring the location of damage to Dunhuang murals based on YOLACT. First, addressing the limitation of traditional YOLACT models in extracting fine-grained texture features, a residual self-attention mechanism capable of extracting global features is introduced to enhance the model's feature extraction capability (mAP50: 79.2064%). Subsequently, by introducing a depth-separable convolutional architecture, the goal of lightweight design is achieved while maintaining high accuracy (parameter count: 715.12 k, floating-point operations: 70.219 M). Based on the above algorithm, we proposed a rapid detection and deployment application method to help Dunhuang mural restoration experts in practical application. Finally, to validate the model's effectiveness, this paper conducted validation on a public dataset. Experimental results show that, compared to three mainstream models, the proposed method achieves the highest mAP50 and the smallest parameter count and floating-point operations.

This study introduces an explainable deep learning model for detecting and classifying dress code violations, leveraging a custom dataset of 130 images categorized into four classes: illegal male dressing, illegal female dressing, legal male dressing, and legal female dressing. The proposed model was built on a pre-trained MobileNetV2 architecture, fine-tuned to achieve a training accuracy of 100% and a validation accuracy of 90%. The model's performance is further validated through a confusion matrix, demonstrating robust classification capabilities, particularly for legal male and female dress codes, with minor misclassifications in the illegal categories. To ensure interpretability, SHAP (SHapley Additive exPlanations) and Gradient Magnitude Heatmaps are employed, providing insights into the model’s decision-making process. The SHAP visualizations reveal the pixel-level contributions to the predictions, while the Gradient Magnitude Heatmaps highlight regions of sensitivity, emphasizing the model's focus on distributed patterns across the images. The alignment between these techniques confirms the reliability of the model's feature extraction capabilities and underscores its generalizability. This paper not only achieves high classification accuracy but also integrates explainability techniques to enhance transparency and trust, making it suitable for socially sensitive applications. The results demonstrated the effectiveness of combining high-performance deep learning models with robust explainability frameworks to address complex classification challenges.

Non-orthogonal multiple access (NOMA) is indeed an important technique of 5G and beyond 5G (B5G) communication systems, as it can increase the spectrum efficiency and system capacity under the same bandwidth. Long-distance unmanned aerial vehicle (UAV) networks, comprising a base station and multiple UAVs, are commonly used for reconnaissance purposes due to the extended flight range of the UAVs, which can provide broad coverage and satisfy the requirements for real-time communication. To alleviate spectrum underutilization of single channel orthogonal multiple access, we investigate resource allocation in UAV networks by utilizing NOMA and then formulate an optimization problem that maximizes the network capacity. The formulated problem can be decomposed into two sub-problems, namely UAV grouping and power allocation optimization. Accordingly, we present a UAV grouping method with low complexity according to the difference of channel gain between UAVs and base station, in which the UAVs with larger difference will be in the same cluster such that they can transmit signals over the same channel. Then, a power allocation optimization algorithm based on Lagrange multiplier method is then used to optimize the transmit power of UAVs utilizing the same channel. This algorithm considers the quality of service requirements for each UAV, and assigns power levels that maximize the overall network capacity. Simulation results have shown that the presented algorithms improve network capacity by approximately 19% compared with that of orthogonal multiple access, which demonstrates the effectiveness of utilizing NOMA in long-distance UAV networks.

Unmanned aerial vehicle (UAV) aerial images often present challenges such as small target sizes, high target density, varied shooting angles, and dynamic poses. Existing target detection algorithms exhibit a noticeable performance decline when confronted with UAV aerial images compared to general scenes. This paper proposes an outstanding small target detection algorithm for UAVs, named Fine-Grained Feature Perception YOLOv8s-P2 (FGFP-YOLOv8s-P2), based on YOLOv8s-P2 architecture. We specialize in improving inspection accuracy while meeting real-time inspection requirements. First, we enhance the targets’ pixel information by utilizing slice-assisted training and inference techniques, thereby reducing missed detections. Then, we propose a feature extraction module with deformable convolutions. Decoupling the learning process of offset and modulation scalar enables better adaptation to variations in the size and shape of diverse targets. In addition, we introduce a large kernel spatial pyramid pooling module. By cascading convolutions, we leverage the advantages of large kernels to flexibly adjust the model’s attention to various regions of high-level feature maps, better adapting to complex visual scenes and circumventing the cost drawbacks associated with large kernels. To match the excellent real-time detection performance of the baseline model, we propose an improved Random FasterNet Block. This block introduces randomness during convolution and captures spatial features of non-linear transformation channels, enriching feature representations and enhancing model efficiency. Extensive experiments and comprehensive evaluations on the VisDrone2019 and DOTA-v1.0 datasets demonstrate the effectiveness of FGFP-YOLOv8s-P2. This achievement provides robust technical support for efficient small target detection by UAVs in complex scenarios.