基于动态特征选择的遥感图像目标检测算法

在遥感图像目标检测领域内，旋转物体的检测存在挑战，卷积神经网络在提取信息时会受制于固定的空间结构，采样点无法聚焦于目标；遥感图像尺度变化大，不同物体需要具有不同尺度感受野的特征映射，具有单一尺度感受野的特征映射无法包含所有有效信息。基于此，提出了可变形对齐卷积，根据候选边框调节采样点，并根据特征映射学习采样点的细微偏移，使采样点聚焦于目标，从而实现动态特征选择；同时提出了基于可变形对齐卷积的感受野自适应模块，对具有不同尺度感受野的特征映射进行融合，自适应地调整神经元的感受野。在公开数据集上的大量实验验证了所提算法可以提高遥感图像目标检测的精度。 

Remote sensing target detection based on dynanic feature selection

In the field of remote sensing image target detection, there still are challenges in oriented object detection. Convolutional neural network is subject to a fixed spatial structure when extracting information, and sampling locations cannot focus on objects. The scale of the remote sensing image varies greatly, and different objects require receptive fields of different scales to obtain feature map. Meanwhile, feature map with a single-scale receptive field cannot contain all effective information. In response to the first problem, deformable alignment convolution was proposed, which can first adjust the sampling locations according to the region of interest, and further learn slight offsets according to feature map, so that sampling locations can focus on objects and realize dynamic feature selection. For the second question, receptive field adaptive module based on deformable alignment convolution was proposed to fuse feature map with receptive fields of different scales and adaptively adjust the receptive field of neurons. Extensive experiments on public datasets showed that this method can improve the accuracy of remote sensing image target detection.