Image Semantic Segmentation Method Based on Deep Learning in UAV Aerial Remote Sensing Image

Abstract

The existing semantic segmentation methods have some shortcomings in feature extraction of remote sensing images. Therefore, an image semantic segmentation method based on deep learning in UAV aerial remote sensing images is proposed. First, original remote sensing images obtained by S185 multirotor UAV are divided into smaller image blocks through sliding window and normalized to provide high-quality image set for subsequent operations. Then, the symmetric encoding-decoding network structure is improved. Bottleneck layer with 1 × 1 convolution is introduced to build ISegNet network model, and pooling index and convolution are used to fuse semantic information and image features. The improved encoding-decoding network gradually strengthens the extraction of details and reduces the number of parameters. Finally, based on ISegNet network, five-classification problem is transformed into five binary classification problems for network training, so as to obtain high-precision image semantic segmentation results. The experimental analysis of the proposed method based on TensorFlow framework shows that the accuracy value reaches 0.901, and the F1 value is not less than 0.83. The overall segmentation effect is better than those of other comparison methods.