Introduction
Detecting aircraft in high-resolution remote sensing imagery is a critical task for civil and military applications. Traditionally, this requires training object detection models like YOLO or Faster R-CNN using thousands of manually annotated bounding boxes. This "fully supervised" approach is labor-intensive and expensive. Researchers have proposed AlexNet-WSL, a weakly supervised learning (WSL) framework that achieves high-precision aircraft detection using only image-level labels (e.g., "aircraft present" vs. "aircraft absent"), significantly reducing the annotation burden.
Methodology: From Classification to Localization
The core of the AlexNet-WSL approach is transforming a standard classification network into a localization tool. The authors modified the classic AlexNet architecture to preserve spatial information that is typically lost in fully connected layers.
The process involves several key steps:
- Architecture Modification: The final fully connected layers of AlexNet are replaced with a Global Average Pooling (GAP) layer and additional convolutional layers. This ensures that the spatial distribution of features is maintained until the end of the pipeline.
- Class Activation Mapping (CAM): By analyzing the weights of the final layer, the model identifies which regions of the feature maps most strongly indicate the presence of an aircraft. This generates a Heat Map representing the object's location.
- Adaptive Binarization: To convert the fuzzy heat map into discrete detections, the system employs Otsu’s method (maximum inter-class variance) to determine an optimal threshold, effectively segmenting the aircraft from the background.
The Weakly Supervised Aircraft Detection Dataset (WSADD)
To validate the model, the researchers created the WSADD benchmark. This dataset consists of 700 images (768x768 pixels) sourced from Google Earth, covering diverse airports globally. It includes 400 positive samples (containing aircraft) and 300 negative samples (runways, aprons, etc.), captured across various seasons and lighting conditions to ensure model robustness.
Performance and Efficiency
AlexNet-WSL demonstrates that high-quality detection does not always require high-quality (expensive) labels. When compared to state-of-the-art fully supervised models, the results are striking:
- Accuracy: AlexNet-WSL reached 97.77% accuracy, nearly matching YOLOv3 (99.16%) and Faster R-CNN (98.88%).
- False Alarm Rate (FAR): The proposed method achieved a lower FAR (0.57%) than both Faster R-CNN (1.39%) and YOLOv3 (0.84%), suggesting it is less prone to background noise.
- Training Speed: The most significant advantage is efficiency. AlexNet-WSL converged in just 10 minutes, whereas Faster R-CNN and YOLOv3 required 3 and 10 hours, respectively, on identical hardware.
Conclusion
The AlexNet-WSL framework proves that spatial information inherent in CNN feature maps can be leveraged for object localization with minimal supervision. While future work is needed to refine bounding box precision and handle extremely dense clusters, this approach offers a highly efficient alternative for large-scale remote sensing tasks where manual annotation is impractical.
Summary updated