Dual-Enhanced-CNN for Few-Shot Object Detection in Remote Sensing Images

Abstract

Existing convolutional neural networks (CNNs) still perform excellently in object detection over remote sensing images, although remote sensing images are more complicated than natural images. But these approaches are very dependent on the quality and quantity of data. when the number of annotated samples gets smaller, the performance of existing CNNs sharply gets worse. Few-shot object detection (FSOD) can alleviate this problem but still has a lot of improvement space. In this work, to further improve the detection performance, We propose our Dual-Enhanced-CNN model. And the main improvements are as follows: 1) We design a weighted cross image attention to learn the interaction information across both images and channels and then improve the detection capabilities of the query image. 2) We design a new adaptive weight loss to focus more on the targets from novel classes and the targets with poor detection performance. We have conducted multiple experiments on the large-scale remote sensing dataset named DIOR. And the higher detection accuracy and relatively stable experimental performance prove the superiority of our method.