Multi-scale Fusion based Multi-stage Small Object Detection in Aerial Images ∗

In aerial images, the objects are mostly small. The number of objects is large and the scale is diverse, so it is difficult to extract the features of multiple scale objects at the same time. The location distribution of object in aerial images is usually dense, making it difficult to locate. These factors bring great challenges to aerial image object feature extraction, and then reduce the performance of detection. Therefore, a multi-scale fusion based multi-stage small object detection method (MSMSD) for aerial images is proposed in this paper. MSMSD adopts EfficientNet as feature extraction backbone and add deformable convolution blocks to achieve better detection performance on objects with irregular shapes. NAS-FPN is leveraged to fuse multi-scale features effectively. A cascade detection mechanism is also designed to reduce noisy detection and mis-detection in this task. In experiment section, the proposed MSMSD outperforms five benchmark object detection algorithms on two aerial image datasets. Experimental results demonstrate that MSMSD can handle the small object detection task in aerial images efficiently.