基于RHTC网络的飞机目标检测与精细识别

飞机目标的方向检测和精细识别是高分辨率光学遥感图像解译领域的一个重要任务。针对遥感图像中多方向密集排布飞机的方向检测和识别困难问题, 提出一种基于旋转混合任务级联(rotated hybrid task cascade, RHTC)网络的飞机检测识别方法。首先, 基于混合任务级联(hybrid task cascade, HTC)网络, 扩展分割分支数量, 并将分割分支与包围框分支多层级联以不断加强语义特征。其次, 设计了一个新的斜框回归器, 将其添加在掩膜分支的最后一层以完成目标方向预测。最后, 增加一个新的方向损失函数以优化训练过程, 从而完成RHTC网络构建。在数据预处理阶段, 构建了每类型号飞机目标的精细掩膜以增强目标细节和提高掩膜预测精度。基于DOTA和公开Google图像构建的飞机数据集开展了多组实验。结果表明, 与其他多种先进的方法相比, 所提方法在飞机检测方向精准度和类别平均精准度上性能更优。此外, 所设计的斜框回归器和方向损失函数在嵌入到其他分割网络时也具有良好的性能。

Aircraft target detection and fine-grained recognition based on RHTC network

Direction detection and fine-gained recognition of aircraft targets is an important task in the field of high-resolution optical remote sensing image interpretation. Aiming at the difficulty of direction detection and recognition of multi-directional densely arranged aircraft in remote sensing images, an aircraft detection and recognition method based on rotating hybrid task cascade (RHTC) network is proposed. Firstly, based on hybrid task cascade (HTC) network, the number of segmented branches is expanded, and the segmented branches and bounding box branches are cascaded at multiple levels to continuously strengthen semantic features. Secondly, a new slant frame regressor is designed and added to the last layer of the mask branch to complete the target direction prediction. Finally, a new directional loss function is added to optimize the training process, so as to complete the construction of RHTC network. In the data preprocessing stage, the fine mask of each type of aircraft target is constructed to enhance the target detail and improve the mask prediction accuracy. Several groups of experiments were carried out on aircraft data sets constructed based on DOTA and public Google images. The results show that compared with other advanced methods, the proposed method has better performance in aircraft detection direction accuracy and category average accuracy. In addition, the designed skew frame regressor and directional loss function also have good performance when embedded into other segmented networks.