To solve the problems of texture lacking and resolution coarseness in the detection of dim and small drone targets in infrared images, we propose a novel RetinaNet with an asymmetric attention fusion mechanism for dim and small drone detection. First, we propose a super-resolution texture-enhancement network as an effective solution for the lack of texture-related information on small infrared targets. The network generates super-resolution images and enhances the texture features of the targets. Second, considering the inadequacy of feature pyramids in the feature fusion stage, we use an asymmetric attention fusion mechanism to constitute an asymmetric attention fusion pyramid network for cross-layer feature fusion in a bidirectional manner; it achieves high-quality semantic and location detail information interaction between scale features. Third, a global average pooling layer is employed to capture global spatial-sensitive information, thus effectively identifying features and achieving classification. Experiments were conducted by using a publicly available infrared image dim-small drone target detection dataset; the results show that the proposed method achieves an AP of 95.43% and a recall of 80.6%, which is a significant improvement over the current mainstream target detection algorithms.
Keywords: RetinaNet; asymmetric attention fusion; drone detection; infrared image; super-resolution.