首页 > , Vol. , Issue () : -
Abstract: Objective:Aircraft detection by deep learning is a hot field in remote sensing image analysis. However, due to the limited perspective of satellite imagery and high similarity in appearance, aircraft type recognization is still a challenging task. The existing deep learning methods cannot be satisfied with the fine-grained aircraft type recognition tasks well, which require a refined lables for datasets. Thus, for recognizing aircraft type recognition in remote sensing images, in this paper, we propose an target segmentation and key point detection integrated aircraft type recognition method. Method: The method organically combines the multi-task deep neural network with the conditional random field and template matching algorithm, and achieves the high-precision recognition of the aircraft type by means of "pre-training, fine-tuning, post-processing". First, based on the multi-task learning and transfer learning technology, the aircraft target position, mask and key point recognition are realized. Secondly, in order to facilitate the high-precision template matching in the later stage, the aircraft target mask refinement algorithm and the key-point based mask attitude adjustment algorithm are proposed to achieve the boundary refinement of the recognition target and the aircraft target mask attitude adjustment. Finally, based on the aircraft type template library constructed in this paper, the refined aircraft mask information is matched with the template library to recognize the aircraft type. Results: The algorithm was applied to the MTARSI data set and remote sensing images for verification. The results showed that the recognition accuracy of 10 types of aircraft could reach more than 85%. Aircraft with simple structure and unique shape have higher recognition accuracy, such as B-2, B-1, etc., while aircraft with complex structure and high similarity with other types of shapes have lower recognition accuracy, such as E-3 reconnaissance aircraft. The algorithm is compared with traditional algorithms and end-to-end deep learning methods, and the results show that 11 types of aircraft are 15.4% and 20.7% more accurate than the two contrasting methods, respectively. Conclusion: The use of target segmentation and key point information has achieved good results in model recognition on high-resolution remote sensing images. However, limitations remain in the breadth of identifiable aircraft types. Therefore, further research is needed.