Rotation-Lossless Non-Local Attention and Fine-Grained Feature Enhancement Based Fine-Grained Oriented Object Detection in Remote Sensing Images

Context and fine-grained features are crucial for fine-grained oriented object detection in remote sensing images. The commonly used non-local attention like self-attention fails in oriented scenes. The similarities of similar objects are low caused by orientation. And fine-grained object detection performance of the existing detectors needs further improvement. In this paper, a novel Rotation-lossless Attention and Fine-grained feature Enhancement based detection network (RAFE) is proposed. Specially, Rotation Lossless Attention (RLA) is designed for suppressing the potential influence of rotation on similarity measurement of similar objects and capturing accurate context information. Then a fine-grained feature enhancement module is proposed to enhance fine-grained features, including De-redundant Feature Fusion (DFF) and Channel Feature Enhancement (CFE). DFF is developed to separate de-redundant fine-grained and context features from dense scale features based on a difference metric function and fuse them. CFE further enhances fine-grained features through the similarity between feature channels. Extensive experiments performed on large-scale remote sensing fine-grained and oriented object detection dataset, FAIR1M, demonstrate that the proposed RAFE achieves competitive performance compared with state-of-the-art methods. The overall accuracy is 42.7% and 5.5% higher than the baseline method, Gliding vertex.