Aircraft State Recognition System Based On Convolutional Neural Network

With the development of science and technology,the automatic recognition technology of aircraft status has become a hot spot and focus of research in the field of computer vision,and is widely used in the field of military and civil aviation.In recent years,algorithmic recognition technology based on deep learning has shown superior performance for image recognition.Based on deep learning algorithm recognition technology.This study proposes a recognition technique through an improved convolutional network.By improving the traditional convolutional network,a convolutional neural network with appropriate depth is built.The improved content includes: convolution depth,activation function,link structure,etc.,and the training set and validation set are used to comprehensively evaluate the recognition ability of the model.Comparing and analyzing the training set and validation set of different depths with the traditional convolutional network training set and test set,the optimal convolutional network network model is obtained,the optimized parameter combination is adjusted,the activation function is improved,the network recognition accuracy is improved,and the generalization ability is strengthened..This paper mainly focuses on how to improve the traditional network link to build a convolutional neural network with high precision,low loss and few parameters.Firstly,on the basis of in-depth study of the link structure of traditional convolutional network links,the traditional convolutional network links are improved,and the links are improved in both horizontal and vertical directions.The network depth is increased in the horizontal direction and batch normalization is added.In order to transform the deep neural network,the modules of the same topology are stacked vertically.Then a convolutional network model is built on the basis of the improved link to extract the salient features of the image.The saliency feature extraction is divided into two steps: network construction and network training.The extracted feature map is used for classification evaluation through the classifier.Experiments are carried out on the basis of the convolutional network model and the training and test results are compared.The main content of comparison includes evaluation indicators,algorithm accuracy and mini-batch loss.The experimental design scheme is as follows: the input image is randomly scaled,histogram equalized,and randomly added with salt and pepper noise to produce 1443 pictures.1298 pictures are randomly selected and divided into 5 categories.The convolution depth of the traditional convolutional network will not exceed 3 layers,and the number of layers of the convolutional network is improved to build 4,5,and 6 different depth networks.Compare the mini-batch loss and mini-batch accuracy of these network outputs.The experimental results show that: in the 5 types of aircraft states with 4-layer network structure,the correct rate of the validation set mini-batch is 91.67%,the training set mini-batch correct rate is93.67%,and the mini-batch loss is 0.4869.Much higher than traditional convolutional networks.Finally,it solves the problems of deepening or widening the number of convolutional network layers,resulting in a huge amount of parameters and weak generalization ability.Use hyperparameter non-convex,non-smooth properties for random combinations to reduce parameter settings.The traditional Re LU activation function is replaced,and the comparison experiment shows that the activation function Leak Re LU has the highest accuracy for network training results,reaching 98.75%,and the mini-batch loss is 0.0322,which is in line with the improvement expectations.