Small Target Detection Algorithm for Air-to-Ground Based on C2Dgnet

This paper deeply explores the small target image detection algorithm for air-to-ground based on C2DGnet, and analyzes its performance and resource consumption through a series of experiments. The focus of this study is the improvement of deep learning models to enhance the efficiency in the task of small target image detection in air-to-ground scenarios. During the research process, we first recognized the significant impact of network architecture and operations on the detection performance and resource consumption of aerial images. By comparing the performance of different variants of network models, we discovered some important findings, namely the potential of CA and Ghost module improvements. These improvements successfully enhanced the detection accuracy while maintaining computational efficiency. Furthermore, the introduction of the CA module further enhanced the feature extraction capability and the performance of small target image detection in air-to-ground scenarios. Ultimately, by improving the Deformable-Transformer model, including the CA module and Ghost module, the detection accuracy was successfully improved while maintaining computational efficiency, with Precision and Recall values of 0.8836 and 0.913 respectively, and an mAP of 0.995. This demonstrates its effectiveness in terms of precision and recall.