A coarse-to-fine small object detection framework based on a background complexity classification strategy

Object detection has achieved great progress and is used in various tasks. However, detecting small objects with lack of appearance information is still a challenging task. It is found that even if the training set with rich images is used to train a network, the small objects in the image with different background complexity cannot be well detected. To address the above issue, this paper proposes a novel small object detection framework based on a background complexity classification strategy specific to the contradiction by adopting the idea of "divide and rule" in philosophy. Firstly, a Background Complexity Classification Network (BCCResNet) is proposed to coarsely classify the input images into three categories according to the complexity of their background textures. Then, a detection network with parallel structure is designed by using mainstream models to detect small objects for three categories of images. Extensive experiments are conducted on two small object detection datasets, i.e., AI-TOD and DOTAv1.0. Our proposed method can significantly improve AP of small object detection, showing effectiveness and advantages.