Optimization Of Deep Learning Model Based On Intelligent Computing

With the rapid development of science and technology,geoscience data grows exponentially,showing the characteristics of big data such as large amount of data,fast update,multi polarization,multi angle and high-speed development.Therefore,how to mine valuable information from these massive geoscience data has become the focus of work in the field of mathematical geoscience in the new era.As a hot research object in the field of machine learning in recent years,deep learning has the ability to automatically extract high-level representation from complex data.It undoubtedly has great potential in solving big data problems such as target detection and remote sensing image recognition.However,deep learning is still in the exploratory stage in geoscience big data mining and analysis.In order to fully tap the potential of deep learning model in geoscience data application,there are still many problems to be solved.In particular,the current deep learning algorithm itself has no standard method to determine the super parameter setting and network architecture design.It almost depends on manual design,which can not completely avoid manual intervention.Moreover,in the context of massive data,the super parameters of deep learning algorithm are as few as dozens and as many as hundreds.The influence of each super parameter on the deep learning model is not independent,and it is inevitable to have deviation in manual setting.Therefore,it is necessary to study the automatic parameter optimization method to truly realize the autonomous,efficient and accurate acquisition of better parameter combination for the deep learning model.Intelligent optimization algorithms(such as genetic algorithm and particle swarm optimization algorithm)have excellent ability to optimize the parameters of complex functions,and can optimize the parameters of deep learning model efficiently and accurately.The work of this thesis mainly focuses on how to use intelligent optimization algorithm to optimize the parameters of deep learning model.The parameter optimization problem of deep learning model is studied from the theoretical level and application level respectively.Experiments show that the algorithm designed in this thesis can effectively obtain a better parameter combination for deep learning model.The main research contents and innovations of this thesis are as follows:(1)A two-stage hyperparametric optimization algorithm based on intelligent optimization method is designed.At present,the determination of parameters in deep learning model mainly depends on manual experience,which is difficult to avoid deviation.To solve this problem,a two-stage optimization algorithm is designed to obtain the deep learning model with the optimal combination of super parameters efficiently and accurately.Considering the efficiency and accuracy,the two-stage convolution neural network optimization algorithm designed in this thesis has a better comprehensive effect on the image classification task based on Cifar-10 and Cifar-100 data sets.For different convolutional neural network models,the performance of single stage and two stage convolutional neural network models obtained by intelligent optimization algorithm in RSP data set is better than the other three traditional convolutional neural networks,which further verifies the practicability of the algorithm proposed in this thesis.(2)An automatic optimization algorithm for the initial weight of deep learning model is designed.The initial point of weight value determines whether the deep learning model converges and the speed of convergence.A good initial weight value can effectively improve the training efficiency of the model.Taking the weight value of the deep learning model as the decision variable,this thesis designs the initial weight optimization algorithm of the deep learning model based on the intelligent optimization method,and automatically searches the training starting point of the deep learning model through the intelligent optimization process,so as to accelerate the convergence of the model.Compared with other weight initialization algorithms,the deep learning model obtained by this algorithm has the best performance in the image classification task based on Cifar-10 and Cifar-100 data sets,and achieves the best accuracy value.On the premise of ensuring the same training accuracy,compared with different weight initialization methods,the initialization weight optimization algorithm proposed in this thesis can achieve stable accuracy,faster convergence and higher efficiency on RSP data set.(3)This thesis designs a super parameter optimization algorithm for retinanet target detection framework.The target detection framework based on deep learning also has the problem of super parameter setting,which will affect the detection accuracy and efficiency.In this thesis,one stage target detection framework retinanet is selected to study the intelligent optimization of the super parameters of retinanet framework.Through comparative experiments,it is verified that the target detection framework optimized by super parameters has better effect in target detection accuracy.(4)A convolution neural network model optimized by intelligent optimization algorithm is designed and applied to the recognition of black cotton soil in remote sensing images.The effectiveness and practicability of the algorithm are further verified by the comparative experiment and the comparative test of the classification effect on the actual remote sensing image samples.