High-Resolution Remote Sensing Image Semantic Segmentation Method Based on Improved Encoder-Decoder Convolutional Neural Network
Résumé
In recent years, Convolutional neural network with encoder-decoder structure is a kind of image semantic segmentation method with high accuracy. However, the characteristics of large amount of parameters and high requirements for computing power restrict its application in the fields of limited computing power and high real-time requirement, such as unmanned driving, road monitoring, remote sensing classification and mobile object detection. To solve the above problems, this thesis firstly designs the dilated convolution combination module, which solves the gridding problems while ensuring large receptive field; then, a double-channel encoder-decoder convolutional neural network is built by using the dilated convolution module combined with the depth separable convolution. Using this network, the parameters and computation of semantic segmentation convolution model of high resolution remote sensing image are greatly reduced while maintaining high segmentation accuracy. Through experiments on GID data sets, and compared with a variety of semantic segmentation methods, this thesis verifies the effectiveness and light-weight of this method.
Origine | Fichiers produits par l'(les) auteur(s) |
---|