飞机目标分类的深度卷积神经网络设计优化

Design and Optimization of Deep Convolutional Neural Network for Aircraft Target Classification

针对使用传统方法和神经网络对飞机目标分类时遇到的准确率低、分类种类少等问题,研究了深度卷积神经网络(DCNN)在飞机目标分类中的可行性。为了匹配模型容量、避免过拟合、提高分类性能等,设计了9层DCNN模型,并使用随机梯度下降优化器进行优化。在数据集中选用6类具有代表性的飞机类型进行实验,提出两种正则化级联方式以防止过拟合并加快模型收敛,最终实现了99.1%的飞机分类准确率,由此说明该DCNN模型在飞机目标分类中的有效性。通过归一化混淆矩阵分析分类结果,给出了每类飞机自分类的准确率。此外,设计了一组对比实验,用经典的AlexNet在同一数据集上进行测试,结果表明,所设计的DCNN的准确率高于AlexNet分类算法95.5%。该模型有效地解决了飞机目标分类精度低的问题,在军事和民航飞机目标的分类研究中具有一定的参考价值和应用前景。

Aiming at the problems of low classification accuracy and less classification types in the classification for aircraft targets by using conventional methods and neural networks,the feasibility of deep convolutional neural network(DCNN)models is studied.To match model capacity,avoid overfitting,and improve classification performance,a nine-layer DCNN model is designed and optimized with stochastic gradient descent optimizer.Six representative types of aircrafts are selected in the dataset,and two regularization cascade methods are proposed to prevent overfitting and speed up the model convergence.Finally,an aircraft classification accuracy of 99.1% is achieved,which demonstrates the effectiveness of the DCNN model in aircraft target classification.By analyzing the classification results of the normalized confusion matrix,the accuracy of the self-classification of each type of aircraft is given.In addition,agroup of comparative experiments are designed to test the same dataset with the classic AlexNet.The results show that the proposed DCNN model is superior to the AlexNet classification algorithm with an accuracy improvement of 95.5%.This model effectively solves the problem of low accuracy in aircraft target classification at present and proves that the DCNN model has certain reference values and application prospects in the classification research of military and civil aviation aircraft targets.