An Efficient High-Resolution Remote Sensing Object Detection Approach Using a Multilevel Optimized RT-DETR Model

Remote sensing images object detection (RSD) involves identifying the position and categorization of objects found in these images. Nevertheless, remote sensing images (RSI) possess characteristics such as small object size, multi-scale variations, and complex backgrounds, which create considerable difficulties for object detection. In recent years, deep learning algorithms have exhibited strong detection abilities in computer vision. However, existing object detection methods do not perform well on RSI. For example, analyzing images with complex backgrounds and dense targets can be quite difficult. This paper proposed an improved version of the real-time detection transformer (RT-DETR) to tackle these problems. In the encoder part, convolutional block attention module (CBAM) and group convolution (GConv) are integrated to optimize convolutional neural network (CNN)-based cross-scale feature fusion module (CCFM), which improved the model’s capability to fuse features across different scales. To extract more detailed information, the S2 shallow feature of ResNet18 is incorporated into the neck network, and subsequently, the shallow feature extraction block (SFEblock) is utilized to extract the shallow feature. Finally, to enrich the diversity of the extracted features, an improved residual block (Resblock) is adopted to extract features from the S4 and S5 feature layers of ResNet18. The experimental results indicate that only 20.09 million parameters are utilized on the NWPU VHR-10 dataset, and the mAP50 of the improved RT-DETR model reaches 91.0%, which is 1.3% higher than that of RT-DETR. Meanwhile, the method’s adaptability and generalization capabilities are further validated on the RSOD dataset, demonstrating its robustness across different scenarios.