Remote Sensing Image Road Segmentation Detection Method Based on Asymmetric Convolution Net

The feature extraction of convolutional layer of neural network has an important influence on the accuracy of neural network identification, and it is very important to increase the ability of neural network to extract image features. The use of suitable convolutional neural network structure in limited practical applications can not be separated from tens of thousands of human operations, which is time-consuming and labor-intensive and easily leads to resource consumption. Thus, it is difficult to improve the performance of convolutional neural network architecture in research. When processing remote sensing images, it is not difficult to find that the road shapes in remote sensing images are often dense and fine, which limits the model to have a certain receptive field. Therefore, on the basis of integrating attention mechanism, this paper adds asymmetric convolution nets as the building blocks of CNN. By manipulating one-dimensional asymmetric convolution nets, the square convolution kernel is enhanced to show its own characteristics, so as to promote the accuracy of network training. That is, symmetric convolution nets are used to replace the original square kernel convolutional layer to construct the asymmetric convolution net (AC-Net). Then AC-Net is replaced by a similar initial architecture to increase the accuracy of the network and avoid unnecessary calculation. The effectiveness of AC-Net is inseparable from its ability to improve the robustness of the model to rotation distortion and the core skeleton of the square convolution kernel. The simulation results demonstrate the feasibility of this research method.