Research On Computing Method Of Distributed Neural Network And Its Application In Aerial Pictures

With the advent of the Internet era,aerial images are widely used for outdoor navigation,green vegetation coverage detection,urban construction area estimation,etc.There is a large demand for images photos and a lot of available information.Because of its high pixel and large amount of data,it is difficult to extract image information.With the development of computer and artificial intelligence,it is more efficient to extract target information from aerial images by computer instead of human.The method of image segmentation(also known as semantic segmentation)in computer vision technology is mainly used to extract the target information in aerial images This method is to mark the coordinate values of different information in the pictures to separate the coordinate value fields of all kinds of information in the aerial images.In this paper,the task of aerial images segmentation is realized by using the neural network.Neural network model and neural network training are the core of image segmentation.The former is mainly to realize the image segmentation algorithm of aerial images,and the latter is mainly to obtain the neural network model.This paper intends to modify the deeplabv3 + neural network model to improve the accuracy of image segmentation,and improve the synchronous training method to achieve the accelerated training of neural network in heterogeneous clusters,and finally realize the image segmentation and application of aerial images.In this paper,we improve the structure of neural network,adjust the super parameters and add the weight loss function to complete the task of image segmentation,and use the expansion prediction to further optimize the image segmentation effect.For the distributed training of neural network,this paper proposes an improved synchronous training method based on heterogeneous cluster.The main idea is to add buffer to complete the queue mechanism of gradient storage,so as to reduce the waiting time between computing nodes and achieve load balancing of clusters,so as to accelerate the training of neural networks.Experiments show that this paper improves the MIOU of image segmentation by improving the model,and improves the training speed of the model by optimizing the cluster.