A Target Detection Algorithm for Remote Sensing Images Based on Deep Learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing Technology

Background:

• Accurate target detection in remote sensing images is crucial for various applications.
• Existing algorithms face challenges with small or complex targets and large image datasets.

Purpose of the Study:

• To develop a novel deep learning algorithm, DFS, for improved remote sensing image target detection.
• To evaluate the performance of DFS against established methods like YOLOv2.

Main Methods:

• Design of a dimension clustering module, a specialized loss function, and sliding window segmentation detection.
• Utilized a large dataset from Google Earth comprising 6 object types: airplanes, boats, warehouses, large ships, bridges, and ports.
• Experimental setup included training, verification, and test sets with 73,490, 22,722, and 2,138 images, respectively.

Main Results:

• DFS demonstrated superior performance, with bridges being the easiest and boats the most challenging targets.
• Achieved a 12.82% improvement in mean average precision (mAP) and a 13% increase in detection accuracy compared to YOLOv2.
• Significantly reduced false positives and increased average Intersection over Union (IOU) by 11.84%.

Conclusions:

• The DFS algorithm offers significant advantages for detecting small targets and processing large remote sensing images.
• DFS provides a more accurate and efficient solution for remote sensing image analysis compared to YOLOv2.
• The algorithm's effectiveness varies based on target size and background complexity, with larger, distinct objects yielding better results.