| In recent years,with its unique advantages,deep learning has been widely concerned in many fields such as computer vision.This paper summarizes machine learning,deep learning and their related algorithms,analyzes the classical neural network model,and introduces the common feature extraction algorithm and data processing technology in detail.In this paper,a Multi-Source Deep Transfer Learning(MS-DTL)classification model is established.Based on the innovation of deep learning algorithm,this model takes full advantages of transfer learning.In order to achieve the goal of multi-task learning in the source domain,the MS-DTL model adopts multi-source training strategy to improve the feature extraction ability of the pre-training model by increasing the diversity of training data.In addition,a composite loss function for different data sets is designed in the source domain for error back propagation.The model adopts a structure based on fully convolutional network,which breaks the limitation of the transferred model on the consistency of data size between source domain and target domain.Hyperspectral images are rich in spectral information,However,due to the insufficient number of labeled samples,when classify the hyperspectral images,a series of problems such as gradient disappearance,gradient explosion and overfitting often occur,especially when the models have great depth.Compared with other larger data sets,the small-scale hyperspectral data sets have more obvious disadvantages in the classification task,and the scarcity of labeled samples is more serious.The experimental results show that in the classification task of small-scale hyperspectral data sets,thanks to the advantages of deep learning network structure,the MS-DTL model can obtain competitive classification accuracy by comparing with other excellent classification algorithms on the premise of greatly saving training time.Thus,the effectiveness of the multi-source deep transfer learning algorithm is further verified. |
