Object detection in remote sensing images using densely connected recursive feature pyramids

In recent years, the multiscale utilization of input sample features has gradually become a research hotspot in the field of target detection. However, remote sensing target detection suffers from some problems, such as small target size, easy confusion with similar objects, and extensive background interference. Therefore, a remote sensing target detection algorithm based on dense connection recursive feature pyramids is proposed. First, the feature fusion mode is improved to use the features of remote sensing images fully. The traditional feature fusion method is only pixel-by-pixel addition, which is simple and rough to calculate and cannot effectively screen features. Therefore, canonical correlation analysis is used to replace the simple pixel-by-pixel additive fusion mode to enhance the effectiveness of feature fusion. Moreover, this method does not add any new parameters. Second, the multireceptive field (MRF) mechanism was added to enhance the feature extraction of small-scale targets, and the features of different receptive fields were extracted and fused by dilated convolution of different sizes to enhance network perception. Given the increase in receptive field types, the richness of features that can be extracted is greatly enhanced, which is conducive to the improved transmission of effective information. In addition, our proposed MRF module is a multibranch convolution module, which is intended to mimic the human visual receptive field mechanism. Then, the feature recurrence form is improved to solve the generalization problem of a multiscale remote sensing target, and the dense connection structure is introduced to enhance the feature fusion density. A dense connection improves network performance because the feature level increases, and the feature richness is enhanced accordingly. Compared with the original recursive feature pyramid, the utilization of the backbone network is remarkably improved. The backbone network and the feature information of high and low layers are fully utilized. Finally,Based on our proposed methods above, this study changes the way of feature recursion and designs a dense connection structure of the recursive feature pyramid. That is, it adds dense connections between multiscale features and each layer of the backbone network to improve the efficiency of feature extraction and utilization. In summary, the network design in this study includes a top-down fusion subnetwork, bottom-up path enhancement subnetwork, and feature recursive fusion subnetwork. Experimental results show that the average accuracy of the proposed pyramid model can be improved by 9.9% on the general dataset MS-COCO2017. On the remote sensing dataset NWPU VHR, the average accuracy of the proposed algorithm can be improved by 1.1%. On the remote sensing dataset DIOR, the average accuracy of the proposed algorithm can be increased by 2.2%, which is higher than other feature pyramid models and detection algorithms. On the large-scale remote sensing dataset DOTA, the average accuracy of the proposed algorithm can be increased by 1.8%. Experimental results show that the proposed method can outperform other feature pyramid models and detection algorithms. It achieves not only high precision detection of remote sensing targets but also has good performance on the benchmark dataset COCO. Therefore, the proposed method is advanced. © 2024 Science Press. All rights reserved.