Towards robust detection of tiny objects in hazy remote sensing images

Object detection in optical remote sensing images (RSIs) is challenging due to cluttered backgrounds, complicated weather conditions as well as variation in the types, scales, and orientations of the objects in question. Numerous methods have been proposed to handle different scenarios in object detection in remote sensing images, but under severe imaging conditions such as hazy scenes and turbulent sensors these methods almost universally fail to perform satisfactorily. In this paper, we propose a new detection framework, called Knowledge-guided Joint Restoration and Detection Network (KJRD-Net), to facilitate the robust detection of tiny objects in hazy optical RSIs. This network is the first method specifically developed to address tiny object detection in hazy RSIs. The network consists of 4 parts: the pre-trained auto-encoder, the parallel image dehazing network and super-resolution network, the image-level coordination block and the object detector. The image is first fed through the auto-encoder to extract comprehensive semantic knowledge. Then, haze removal and super-resolution is performed on the image by the parallel dehazing and super-resolution network. After that, the semantic information and the image restoration cues extracted by the first two components is fed through the image-level coordination block to acquire the final restored image. Finally, the processed images, incorporated with knowledge extracted by prior network structures, are fed into an object detection network. The network is trained end-to-end. In addition, we construct the DOTA-H dataset, a novel dataset focusing on tiny object detection in hazy RSIs. Extensive experiments show that our framework is effective at detecting tiny objects from RSIs of low quality, improving the mAP of the Faster R-CNN baseline on our DOTA-H dataset from 41.2 to 48.6.