关键词: 卷积神经网络, 分层跳连融合, 自适应感兴趣区域提取, 多目标检测分割, 局部融合全连接卷积网络

Abstract:

Aiming at the problems of difficult feature extraction, low accuracy, and high false rate of high-resolution remote sensing images in target detection and segmentation, an improved Mask R-CNN convolutional neural network is proposed. This network uses ResNet50 as a feature extraction network. Based on this, it uses bottom-up and top-down two layer-hop fusion methods to perform better image feature extraction. Aiming at the problem that the size difference between different targets in the remote sensing image is too large and the target is easy to lose, an adaptive region of interest is designed to extract the region of interest. In the target segmentation, the locally fully convolved neural network is used to replace the original full convolutional neural network, and the up-sampling operation is used to replace the de-convolution operation. The verification on the NWPU VHR-10 data set shows that the method significantly improves the accuracy of multi-target detection and segmentation in remote sensing images compared with existing methods.

Key words: convolutional neural network, hierarchical hopping fusion, adaptive region of interest extraction, multi-target detection segmentation, local fusion fully connected convolutional neural network