A Vehicle Detection Method Based on an Improved U-YOLO Network for High-Resolution Remote-Sensing Images

Dudu Guo, Yan Wang, Shunying Zhu, Xin LiSustainability 2023 · DOI

TL;DR

This paper introduces U-YOLO, an enhanced YOLOv3-based network for detecting vehicles in high-resolution remote-sensing images. By integrating adaptive feature fusion, cross-scale channel attention, and sub-pixel convolution, U-YOLO significantly improves detection accuracy, addressing the challenges posed by limited vehicle features and pixel resolution in such imagery.

What Problem Does This Paper Address?

Detecting vehicles in high-resolution remote-sensing images presents a significant challenge due to the inherent limitations of the data. Vehicles often appear as small objects with insufficient feature information and a limited number of pixels, making them difficult to distinguish from the background. Traditional vehicle detection methods struggle with these issues, leading to low accuracy. While deep learning approaches have shown promise, there's a need for specialized architectures that can effectively capture and enhance the subtle features of vehicles in these challenging datasets. This paper aims to bridge this gap by proposing an improved YOLOv3 model tailored for this specific task.

Methods

The core of this work is an improved YOLOv3 architecture, dubbed U-YOLO, which incorporates several novel modules to enhance vehicle detection:

  • Adaptive Fusion (AF) Module: Integrated into the backbone of the YOLO model, this module aims to preserve and enhance the underlying structural information of feature maps. It fuses multi-scale information through a two-layer convolutional structure, reducing the loss of contextual information crucial for localization.
  • Cross-Scale Channel Attention (CSCA) Module: This module is applied during feature fusion. It generates fusion weights to capture the saliency information of vehicles across different scales. By enhancing vehicle features and suppressing background noise, it improves the distinguishability of targets, especially in images with varying target scales.
  • Sub-Pixel Convolution (SC) Module: Replacing the traditional linear interpolation for up-sampling, the SC module reorganizes channel information within the feature layer. This process amplifies the feature map resolution, thereby enhancing the integrity and detail of small vehicle features and improving overall detection accuracy.

Datasets

The proposed U-YOLO model was trained and evaluated on two publicly available remote-sensing image datasets:

  • NWPU VHR-10: This dataset contains 800 aerial remote-sensing images with 10 target types.
  • DOTA: A larger dataset comprising 2806 remote-sensing images, also featuring various object categories.

The authors also utilized a custom dataset, V547, for further validation, which consists of 547 remote-sensing images with approximately 5000 vehicle samples, derived from sources like Google Earth and Worldview-3.

Key Insights and Ideas

  • The Synergy of Modules: The paper demonstrates that combining AF, CSCA, and SC modules is crucial for achieving significant performance gains. Each module contributes uniquely to enhancing feature representation, context, and resolution.
  • Attention for Saliency: The CSCA module's ability to generate cross-scale channel attention weights is key to highlighting vehicle features and reducing background interference, which is particularly effective for small and subtle targets.
  • Sub-pixel for Detail: Replacing linear interpolation with sub-pixel convolution proves effective in up-sampling feature maps without losing critical details, which is vital for small object detection.
  • Backbone Matters, but Enhancements are Key: While different backbone networks (Resnet50, Resnet101, Efficient-B2, Darknet53) were tested, the proposed U-YOLO enhancements consistently improved performance across these backbones, indicating the robustness of the core U-YOLO architecture.

Key Results

  • NWPU VHR-10: U-YOLO achieved an mAP of 91.35% with the Darknet53 backbone, an improvement of 6.89% over the original YOLOv3.
  • DOTA: U-YOLO achieved an mAP of 71.38%, an improvement of 4.94% over the original YOLOv3.
  • Comparison with Other Networks: U-YOLO outperformed RFBNet, M2det, and SSD300 on the V547 dataset, with AP values improved by 6.84%, 6.38%, and 12.41%, respectively. On a real-road remote-sensing image test, U-YOLO achieved 94.5% accuracy with a 6.8% false detection rate, outperforming the compared methods.
  • Module Contributions: The CSCA module was found to contribute the most to the accuracy improvement, followed by the AF module, and then the SC module.

Limitations and Future Work

  • Shadow Occlusion: The authors acknowledge that the proposed method struggles with dark vehicles in shadowy areas, leading to missed detections. This is attributed to amplified contrast between dark vehicles and the background.
  • Future Work: The researchers plan to address the shadow occlusion issue by exploring methods to enhance sparse features of dark vehicles and extract global information to mitigate detection accuracy degradation caused by shadows.

Summary updated

Abstract

The lack of vehicle feature information and the limited number of pixels in high-definition remote-sensing images causes difficulties in vehicle detection. This paper proposes U-YOLO, a vehicle detection method that integrates multi-scale features, attention mechanisms, and sub-pixel convolution. The adaptive fusion module (AF) is added to the backbone of the YOLO detection model to increase the underlying structural information of the feature map. Cross-scale channel attention (CSCA) is introduced to the feature fusion part to obtain the vehicle’s explicit semantic information and further refine the feature map. The sub-pixel convolution module (SC) is used to replace the linear interpolation up-sampling of the original model, and the vehicle target feature map is enlarged to further improve the vehicle detection accuracy. The detection accuracies on the open-source datasets NWPU VHR-10 and DOTA were 91.35% and 71.38%. Compared with the original network model, the detection accuracy on these two datasets was increased by 6.89% and 4.94%, respectively. Compared with the classic target detection networks commonly used in RFBnet, M2det, and SSD300, the average accuracy rate values increased by 6.84%, 6.38%, and 12.41%, respectively. The proposed method effectively solves the problem of low vehicle detection accuracy. It provides an effective basis for promoting the application of high-definition remote-sensing images in traffic target detection and traffic flow parameter detection.
BibTeX
@inproceedings{guo2023vehicle,
  author = {Dudu Guo and Yan Wang and Shunying Zhu and Xin Li},
  title = {{A Vehicle Detection Method Based on an Improved U-YOLO Network for High-Resolution Remote-Sensing Images}},
  booktitle = {Sustainability},
  year = {2023},
  doi = {10.3390/su151310397},
}
Figure