Multi-scale Object Detection of Remote Sensing Images Based on Efficient Feature Extraction and Transformer

Object detection in remote sensing images is a hot research topic in the field of geosciences and remote sensing. In recent years, many studies have successfully used Convolutional Neural Networks(CNN) to improve the performance of object detection in remote sensing images. CNN efficiently captures high-level semantic features through convolution and pooling operations, but occasionally loses detailed features related to small objects, which cannot solve the multi-scale problem of remote sensing images well. In addition, CNN is conducive to the extraction of local features, there are some defects in the grasp of global information. However, Transformer based on attention mechanism can obtain long-distance RSI relationship. Therefore, this paper studies the remote sensing object detection Transformer. Specifically, The EFET framework proposed in this paper is a combination of CNN and Transformer. CNN is used to capture the multi-scale features of the detected objects, and Transformer is used to encode and decode the extracted features to complete the object detection task of remote sensing images. We output the feature maps of each stage of the backbone network through Cross-Link to enhance the modeling of object information at different scales. Considering the information loss caused by the 1x1 convolution operation on the channel dimension of the feature map, the channel attention is introduced to capture the inter-channel dependence, the recognition of small targets has a good detection effect. For the over-fitting problem of remote sensing image datasets, data enhancement is combined with the model to improve the application ability of the model. A large number of experiments on NWPU VHR-10 and DIOR object detection datasets verify the effectiveness of our method.