RADet: Refine Feature Pyramid Network and Multi-Layer Attention Network for Arbitrary-Oriented Object Detection of Remote Sensing Images

Abstract

Object detection has made significant progress in many real-world scenes. Despite this remarkable progress, the common use case of detection in remote sensing images remains challenging even for leading object detectors, due to the complex background, objects with arbitrary orientation, and large difference in scale of objects. In this paper, we propose a novel rotation detector for remote sensing images, mainly inspired by Mask R-CNN, namely RADet. RADet can obtain the rotation bounding box of objects with shape mask predicted by the mask branch, which is a novel, simple and effective way to get the rotation bounding box of objects. Specifically, a refine feature pyramid network is devised with an improved building block constructing top-down feature maps, to solve the problem of large difference in scales. Meanwhile, the position attention network and the channel attention network are jointly explored by modeling the spatial position dependence between global pixels and highlighting the object feature, for detecting small object surrounded by complex background. Extensive experiments on two remote sensing public datasets, DOTA and NWPUVHR -10, show our method to outperform existing leading object detectors in remote sensing field.Remote Sens. 2020, 12, 389 2 of 20 for each region of interest (RoI). Then, the features are used to achieve category-specific classification and regression for the corresponding proposals. Finally, the final detection result is obtained through post-processing, such as non-maximum suppression. Faster R-CNN is a classical two-stage object detector, which is composed of Region Proposal Networks (RPN) and a detection network consists of classifiers and regressors, and can detect objects quickly and accurately in an end-to-end manner. Based on Faster R-CNN, more improved two-stage object detectors such as Region-based Fully Convolutional Networks (R-FCN) [7] and Mask R-CNN [8] were proposed. To further improve the efficiency of the object detector, Joseph Redmon et al. proposed a single stage target detector based on regression, called YOLO [9]. For the simple structure, You Only Look Once (YOLO) is extremely fast, but its accuracy is lower than that of the two-stage detector. Based on YOLO, YOLO v3 [10] and YOLO 9000 [11] were proposed successively. To trade off the detection speed and accuracy, Single Shot MultiBox Detector (SSD) [12] was proposed, whose speed and accuracy were between YOLO series algorithm and R-CNN series algorithm.

Multi-Scale Feature Integrated Attention-Based Rotation Network for Object Detection in VHR Aerial Images

“…For HRSC2016 evaluation, the results are reported by the standard VOC AP metrics with Intersection Over Union (IoU) threshold of 0.5. RoI-Transformer [31], SCRDet [33] and RADet [34] are selected as the representation in three categories of oriented object detections to make a comparison with our proposed MFIAR-Net. As the Table 2 shows, these methods achieve competitive results.…”

“…The channel of our proposed MFIAR-Net is 256 after FPN, which can save much time without performance loss. RADet [34] is based on Mask RCNN [16], which prediction of mask branch is a pixel-to-pixel task. It needs more computation time with 0.24 s speed.…”

“…Table 1 summarizes the experimental results of different methods. As can be seen in Table 1, we compare with FR-O [35], R-DFPN [29], R 2 CNN [32], RRPN [44], ICN [30], RoI-Transformer [31] and recently proposed SCRDet [33] and RADet [34]. The results reported here were obtained by submitting our predictions to the official DOTA evaluation server.…”

“…SCRDet can achieve precisely rotation detection especially for small and cluttered objects on VHR remote sensing images. For the third method, Li et al [34] proposed a rotation detector named RADet inspired by Mask RCNN [16], which obtains the rotated bounding box by calculating the minimum aera rectangle from the correspondingly precited mask shape. This simple presentation with an efficient multi-scale network achieve the competent performance on two benchmarks DOTA [35] and NWPU VHR-10 [36].…”

“…The proposed framework is composed of three modules: Multi-scale Feature Integration Network (MFIN), Double-Path Feature Attention Network (DPFAN) and Rotation Detection Network. Compared with advanced rotation detection methods such as RADet [34], SCRDet [33], RoI-Transformer [31] and ICN [30], our framework is more suitable for multi-class and arbitrary-oriented object detection in aerial images. The main contributions of this paper are as follows:…”