YOLOv8-MPEB small target detection algorithm based on UAV images

Wenyuan Xu, Chuang Cui, Yongcheng Ji, Xiang Li, Shuai LiHeliyon 2024 · DOI

TL;DR

This paper introduces YOLOv8-MPEB, a novel algorithm for detecting small targets in Unmanned Aerial Vehicle (UAV) imagery, specifically addressing challenges like scale variation and complex scenes. By integrating a lightweight backbone (MobileNetV3), an efficient attention mechanism (EMA), a dedicated small target detection layer, and an enhanced feature fusion network (BiFPN), the proposed model achieves a significant improvement in accuracy and a reduction in model size and computational cost, making it more suitable for real-time applications.

What Problem Does This Paper Address?

Detecting small objects in images captured by Unmanned Aerial Vehicles (UAVs) is a critical but challenging task. UAV imagery often features drastic changes in scale, small target sizes, cluttered backgrounds, and varying environmental conditions, leading to missed detections and false positives. This is particularly problematic in applications like construction site safety monitoring, where accurately identifying workers wearing essential safety gear (helmets and reflective clothing) is crucial for preventing accidents. Existing deep learning models, while powerful, can be computationally expensive and struggle with the unique characteristics of UAV data. This paper aims to bridge this gap by developing a more efficient and accurate small target detection algorithm tailored for UAV applications.

Methods

The researchers propose an enhanced YOLOv8 model, dubbed YOLOv8-MPEB, built upon several key modifications:

  • Lightweight Backbone: The original CSPDarknet53 backbone is replaced with MobileNetV3. This significantly reduces the number of model parameters and computational complexity, leading to faster inference speeds, which is vital for real-time UAV applications.
  • Efficient Multi-Scale Attention (EMA) Mechanism: The EMA mechanism is integrated into the Convolution to Feature (C2f) modules. This attention mechanism helps the model focus on salient features and suppress irrelevant ones, improving the extraction of vital information, especially for small targets.
  • Dedicated Small Target Detection Layer: A specialized detection layer is introduced to optimize feature extraction for multi-scale targets, specifically addressing the challenges posed by small objects.
  • Bidirectional Feature Pyramid Network (BiFPN): The Neck of the network utilizes a BiFPN structure. This enhances the fusion of multi-scale features, improving the model's ability to handle scale variations and complex scenes, thereby boosting generalization.
  • P2 Detection Head: An additional P2 detection head is added to the network to specifically improve the detection of small targets, complementing the existing detection layers.

Datasets

The study utilizes a custom-made dataset of helmet and reflective clothing images collected from real-world construction sites. This dataset comprises 2672 images, which were augmented with noise, random flipping, and enhanced brightness to improve model robustness. The final dataset used for training and validation consists of 6680 images categorized into four groups: head, helmet, reflective clothing, and other clothing. The dataset is split into an 8:2 ratio for training and validation.

Key Insights and Ideas

  • Lightweighting for Practicality: Replacing the heavy backbone with MobileNetV3 is a crucial step towards making advanced object detection models deployable on resource-constrained platforms like edge devices, which are common for UAVs.
  • Attention is Key for Small Targets: The integration of the EMA attention mechanism proves effective in guiding the model to focus on subtle features of small targets, a common bottleneck in existing methods.
  • Multi-Scale Fusion is Essential: The BiFPN's ability to effectively fuse features from different scales is critical for handling the drastic size variations inherent in UAV imagery.
  • Specialized Detection for Small Objects: The addition of a dedicated small target detection layer and P2 head directly addresses the core challenge, showing that targeted architectural modifications can yield significant performance gains.

Key Results

The YOLOv8-MPEB algorithm achieved a mean Average Precision (mAP) of 91.9% on the custom helmet and reflective clothing dataset. Compared to the standard YOLOv8 model, this represents:

  • A 2.2 percentage point increase in average accuracy.
  • A 34% reduction in model parameters.
  • A 32% reduction in model size (down to 14.5 MB).
  • A parameter count of 7.39 million.

The paper also highlights that YOLOv8-MPEB outperforms other prevalent detection algorithms in terms of both accuracy and speed. Ablation studies confirmed that each component (MobileNetV3 backbone, P2 detection head, EMA attention, and BiFPN) contributed positively to the overall performance, with the combined YOLOv8-MPEB model achieving the highest mAP.

Limitations and Future Work

The authors acknowledge that challenges remain in detecting tiny, aggregated, and similar targets, which can still lead to missed or false detections. Future work will focus on:

  • Optimizing the multi-scale feature pyramid strategy.
  • Improving the localization loss function.
  • Enhancing algorithm accuracy and model performance in scenarios with small target aggregations.

Summary updated

Abstract

Target detection in Unmanned Aerial Vehicle (UAV) aerial images has gained significance within UAV application scenarios. However, UAV aerial images present challenges, including large-scale changes, small target sizes, complex scenes, and variable external factors, resulting in missed or false detections. This study proposes an algorithm for small target detection in UAV images based on an enhanced YOLOv8 model termed YOLOv8-MPEB. Firstly, the Cross Stage Partial Darknet53 (CSPDarknet53) backbone network is substituted with the lightweight MobileNetV3 backbone network, consequently reducing model parameters and computational complexity, while also enhancing inference speed. Secondly, a dedicated small target detection layer is intricately designed to optimize feature extraction for multi-scale targets. Thirdly, the integration of the Efficient Multi-Scale Attention (EMA) mechanism within the Convolution to Feature (C2f) module aims to enhance the extraction of vital features and suppress superfluous ones. Lastly, the utilization of a bidirectional feature pyramid network (BiFPN) in the Neck segment serves to ameliorate detection errors stemming from scale variations and complex scenes, thereby augmenting model generalization. The study provides a thorough examination by conducting ablation experiments and comparing the results with alternative algorithms to substantiate the enhanced effectiveness of the proposed algorithm, with a particular focus on detection performance. The experimental outcomes illustrate that with a parameter count of 7.39 M and a model size of 14.5 MB, the algorithm attains a mean Average Precision (mAP) of 91.9 % on the custom-made helmet and reflective clothing dataset. In comparison to standard YOLOv8 models, this algorithm elevates average accuracy by 2.2 percentage points, reduces model parameters by 34 %, and diminishes model size by 32 %. It outperforms other prevalent detection algorithms in terms of accuracy and speed.
BibTeX
@inproceedings{xu2024yolov,
  author = {Wenyuan Xu and Chuang Cui and Yongcheng Ji and Xiang Li and Shuai Li},
  title = {{YOLOv8-MPEB small target detection algorithm based on UAV images}},
  booktitle = {Heliyon},
  year = {2024},
  doi = {10.1016/j.heliyon.2024.e29501},
}
Figure