Research And Implementation Of Small Sample Aerial Scene Image Classification Based On Transfer Learning

As a typical depth learning model,Convolution Neural Network(CNN)is a kind of feedforward neural network with convolution calculation and depth structure.With the continuous progress of deep learning technology,CNN makes a great breakthrough in the field of image classification.As a new technology,UAV aerial photography has been widely used in military reconnaissance,map remote sensing and traffic control.It is of great academic research value to use CNN to classify aerial images.However,there are two problems in the classification of aerial images by CNN.On the one hand,the quality of aerial images will be affected by weather factors.Such as low illuminance in cloudy days and low contrast in foggy days,which may bring about low accuracy of image classification.On the other hand,traditional deep learning algorithm needs big data set to model in training,and consumes a lot of time and calculation power as well.While most of aerial data sets are too small to train the traditional machine learning network directly,which may cause serious over-fitting problems.In addition,the generalization ability of the training model will be poor.In order to solve the above two problems,under the background of small sample aerial scene image classification based on transfer learning,tHSI dissertation optimizes from two aspects which includes aerial image enhancement and recognition model improvement.The main contents of tHSI thesis are as follows.(1)In order to solve the problem of low illuminance and low contrast of aerial image in bad weather,which may reduce accuracy of image recognition,tHSI dissertation proposed an improved image enhancement algorithm based on Retinex in HueSaturation-Intensity(HSI)color space.The improved image enhancement algorithm ensures that the original image color will not be distorted,improves illuminance and contrast of aerial image,and avoids image information loss in traditional Retinex algorithm.(2)In order to solve the problem of overfitting while using deep learning to train models with aerial photography small datasets,tHSI dissertation proposed transferred RBM-CNN(tr-RCNN)model,a transfer hybrid model based on CNN and Restricted Boltzmann Machine(RBM).Tr-RCNN combines the feature learning capabilities of the two models to improve the accuracy of aerial image classification under a small data set.(3)THSI dissertation implements a UAV aerial scene recognition system based on transfer learning,which includes software implementation of improved Retinex image enhancement algorithm,image classification algorithm based on tr-RCNN model and Human Machine Interaction(HMI).In addition,the feasibility and application value of tHSI paper has been verified.Through software experiment,information entropy,local contrast and availability of aerial image have been significantly improved while using improved Retinex algorithm.The test results on the open dataset NWPU-RESISC45 show that tr-RCNN model has high image classification accuracy under the training of small target sets.