Abstract: This exploration primarily aims to jointly apply the local FCN (fully convolution neural network) and YOLO-v5 (You Only Look Once-v5) to the detection of small targets in remote sensing images. Firstly, the application effects of R-CNN (Region-Convolutional Neural Network), FRCN (Fast Region-Convolutional Neural Network), and R-FCN (Region-Based-Fully Convolutional Network) in image feature extraction are analyzed after introducing the relevant region proposal network. Secondly, YOLO-v5 algorithm is established on the basis of YOLO algorithm. Besides, the multi-scale anchor mechanism of Faster R-CNN is utilized to improve the detection ability of YOLO-v5 algorithm for small targets in the image in the process of image detection, and realize the high adaptability of YOLO-v5 algorithm to different sizes of images. Finally, the proposed detection method YOLO-v5 algorithm + R-FCN is compared with other algorithms in NWPU VHR-10 data set and Vaihingen data set. The experimental results show that the YOLO-v5 + R-FCN detection method has the optimal detection ability among many algorithms, especially for small targets in remote sensing images such as tennis courts, vehicles, and storage tanks.
Note on Status: The provided document images contain a prominent "RETRACTED" watermark. While the summary below reflects the technical content of the paper as written, readers should exercise caution and consult the publisher's retraction notice for details regarding the validity of the findings.
Introduction
In remote sensing, the "small target problem" is a major bottleneck. Standard object detection models are often optimized for "natural" images where the subject occupies a large portion of the frame. In contrast, a single satellite image might cover several square kilometers, rendering critical objects like vehicles, aircraft, or storage tanks as tiny, densely packed clusters.
Existing models like standard YOLOv1 or SSD (Single Shot MultiBox Detector) often struggle with these small targets because their feature extraction process loses fine-grained spatial information. Conversely, high-accuracy models like Faster R-CNN can be computationally expensive. This research seeks a middle ground: a model that is deep enough to extract subtle features but structured to maintain spatial awareness for tiny objects.
Methodology
The proposed framework integrates the YOLOv5 algorithm with a Region-based Fully Convolutional Network (R-FCN). The technical novelty lies in three specific areas:
1. Hybrid Architecture
The system utilizes YOLOv5 as the foundational detector but augments the pipeline with R-FCN logic. R-FCN is unique because it uses "position-sensitive score maps" to address the dilemma between translation invariance (needed for classification) and translation variance (needed for pinpointing an object's location).
The authors analyzed the evolution of these networks, from Fast R-CNN (FRCN) structures to more advanced fully convolutional approaches to optimize feature extraction.
2. Multi-scale Anchor Mechanism
Borrowing from Faster R-CNN, a multi-scale anchor system was implemented. This allows the model to "look" for objects using various bounding box sizes and aspect ratios simultaneously, making it much more adaptable to targets of different scales.
3. Local FCN Integration
The use of a local Fully Convolutional Network helps in extracting deeper characteristics of image targets without the information loss typically associated with fully connected layers.
Experimental Setup
The model was trained and evaluated on two primary benchmarks:
- NWPU VHR-10: A multi-class geospatial object detection dataset containing 650 high-resolution optical images with 10 target classes, including airplanes, storage tanks, and vehicles.
- Vaihingen Dataset: Specifically used for its high-resolution infrared imagery (0.08-meter resolution) to test the model's robustness across different sensor types.
Results and Discussion
The proposed YOLOv5 + R-FCN method was compared against several baselines, including FRCN-ZF, FRCN-VGG, SSD, and YOLOv1.
- Optimal Detection: The hybrid method achieved the highest recall rates across most categories, particularly for the smallest targets like vehicles and storage tanks.
- Precision-Recall Performance: In the NWPU VHR-10 dataset, the model maintained high accuracy (between 1.0 and 1.2 in the authors' normalized metrics) even when the recall rate was pushed to 0.8.
- Spatial Awareness: The authors argue that the primary reason YOLOv5 struggles with small targets is the lack of detailed grid division. By integrating R-FCN, the model maintains the characteristics of candidate regions more effectively throughout the network.
- Deeper Feature Extraction: The deeper network architecture allowed for better feature extraction in "densely arranged" scenarios, such as parking lots, where individual objects often blur together in shallower models.
Conclusion
The YOLOv5 + R-FCN detection method demonstrates superior performance for small targets in remote sensing images, achieving more accurate feature recognition for densely arranged targets. However, limitations remain regarding data diversity and dependency on large training sets. Future work will focus on redesigning the network structure to reduce data dependence and improve performance in unbalanced conditions where certain target classes are rare.
Summary updated