Lightweight Remote Sensing Image Small Object Detection Based on REM-YOLO

Satellite remote sensing imaging technology captures surface information from Earth’s orbit using sensors mounted on satellites. These images record the radiation energy reflected or emitted by the Earth’s surface across different electromagnetic bands (such as visible light, infrared, and microwave) and convert it into image data for the detection and analysis of geophysical phenomena. Although satellite remote sensing images have been widely used in various studies, existing methods in the field of remote sensing object detection still face significant challenges. The primary reasons include the long distance between the satellite and Earth, resulting in smaller targets within the captured images, which typically have low resolution. Additionally, the complex backgrounds and scarcity of detailed information present further difficulties. When deploying algorithms on ground devices for real-time processing, the challenge of optimizing accuracy and speed under limited computational resources is particularly pressing. To address the aforementioned challenges, this paper proposes an efficient REM-YOLO detector that integrates feature enhancement and multi-scale feature fusion techniques to improve the accuracy and efficiency of small object detection in remote sensing images. To strengthen the neural network’s ability to represent input data features, we introduce a Feature Enhancement algorithm. This algorithm optimizes input features, enabling the model to more accurately capture critical information in the data, thereby enhancing overall object detection performance. Feature enhancement techniques are widely applied in the fields of computer vision and deep learning, providing richer and more reliable data support for subsequent object detection models. In terms of feature processing, we adopt a multi-scale feature fusion technique to enhance the model’s understanding and representation of input data. Feature fusion leverages information from different feature levels: shallow features retain more details such as edges and textures, while deep features contain higher-level semantic information. By integrating these features through a Progressive Feature Pyramid Network, the model can simultaneously utilize multi-level feature information, achieving effective cross-layer connections and context information weighting, thereby improving object recognition and classification capabilities. Furthermore, we incorporate the Soft-NMS re-scoring function, which significantly enhances the performance of NMS by attenuating the scores of overlapping detection boxes that do not cover the target object. This method greatly reduces the false positive rate of detection results, optimizing the network model and achieving more efficient remote sensing object detection. Our approach has demonstrated outstanding results on the DOTA, and DIOR datasets, reducing the number of parameters while maintaining high accuracy and improving detection speed, thereby validating the effectiveness of our algorithm.