Research On Landform Image Classification Algorithm Based On Deep Network

The natural landform image classification of UAV is the basis for UAV autonomous landing in unknown areas and is a research hotspot in the field.However,natural scene images have the characteristics of high similarity between classes,complex scenes,and rich texture information,and are easily affected by factors such as illumination and weather conditions.Traditional image classification algorithms have poor classification results.Therefore,how to improve the performance of natural landform scene classification algorithm has extremely important research significance.In recent years,deep learning has become a research hotspot in the field of machine learning.It has been widely used in the image classification field.The deep network structure is more suitable for extracting the features of the image,and the highlevel semantic features are obtained through layer-by-layer feature learning,which is particularly suitable for handling the classification problem of the UAV natural landform scene.Therefore,based on the deep learning method,this paper studies the natural landform classification of UAV and lays the foundation for the autonomous landing of UAV,aiming at the characteristics of high similarity,complex scene and easy interference between the classes of natural landform scenes.The main contents of this paper are as follows:(1)To overcome the problem of high misclassification rate due to the similarity between classes in landform scene images,a similar landform image classification algorithm based on SAE and SVM is proposed.Firstly,DSEA is constructed to perform feature learning on the image.Then feature reorganized method is used to obtain each layer of adjustment factors and reorganized feature sets of each layer.Finally,with regard to the multi-classification of UAV landform images,the genetic algorithm is introduced into the training of the SVM classifier.Experiments show that the DSAESVM(GA)algorithm can automatically learn the deep complex and abstract features of the image,and the reorganized feature effectively improves the classification accuracy of the similar landform image.(2)To overcome the problem of low scene classification rate and CNN training time is long for complex scenes and rich texture information,an adaptive UAV landform scene classification based on DCT and CNN is proposed.Firstly,the DCTCNN algorithm model is constructed based on the advantages of CNN and DCT.Then according to the characteristics of DCT energy concentration,an adaptive DCT coefficient selection method is proposed to reduce the redundant information of the input data without losing the original input information.Finally,the DCT coefficients adapted to the selection are input into the DCT-CNN model,and the characteristic expressions and the final classification results of the images are obtained through feature learning.Experiments show that the algorithm reduces the data redundancy and training time.Otherwise,the proposed method can automatically learns high-level semantic features and the extracted features have better feature expressions,which improves the classification accuracy of UAV landform scenes in complex environments.(3)To overcome the problem of CNN is easy to fall into a local optimum when the sample size is small,a SC pre-trained convolutional neural network landing point landform classification algorithm is proposed.Firstly,the training samples are subjected to non-subsampled Contourlet transform,and the first two layers of decomposition images are selected to expand the training samples.Then the images are randomly selected using the SC algorithm to learn its local features,and the features are sorted according to the gray mean gradient from the largest to the smallest.Finally,the eigenvalues with large gray mean gradient are selected to initialize the CNN convolution kernel.Experimental results show that using the SC algorithm to learn the characteristics of the original image with statistical characteristics initialize the CNN convolution kernel,obtain better classification effect than traditional underlying visual features,which effectively avoiding the network training from falling into a local optimum.Otherwise,comprehensive high and low frequency with the recognition advantages of different landform scenes,under the condition that the training sample is limited,the classification accuracy of the landing landform of the UAV in the natural landscape scene is improved.