Research On Classification Methods Of Turbulence-degraded Image Based On SVM And CNN

Turbulence image restoration is a research hotspot in the field of degraded images.The research on the restoration of turbulence-degraded images currently mainly stays in the traditional classical image restoration method.The restoration effect of turbulence-degraded images is greatly affected by the fuzzy scale factor k value in the turbulence model.The image restoration effect is not good when the k-value is unknown,but there are few research methods for solving the k-value at the present stage.In this paper,two methods for classifying four different k-value range images are proposed and studied.Firstly,this paper proposes a DT-SVM turbulence-degraded image classification method based on DCT high-frequency features.The method extracts the image features of DCT transformed images of turbulence-degraded images,and further extracts the high-frequency features of DCT-transformed images.The PCA dimensionality reduction technique combined with DT-SVM multi-classification method is used to realize the classification experiments of turbulent-degraded images with different k-value ranges.Secondly,this paper proposes a turbulence-degraded image classification method based on improved CNN network model.The method is based on the powerful feature abstraction extraction and automatic learning classification ability of convolutional neural networks.Experiments are carried out under the classical LeNet-5 network model.It is found that the model can not converge well.Based on experimental data,the LeNet-5 convolutional neural network model is studied.Structure and parameters,and proposed improvements on the basis of this network structure,and used the improved network to classify the turbulence-degraded images with different k-value ranges;The experimental results show that the overall classification accuracy is close to 60% on the DT-SVM turbulence-degraded image classification method based on DCT high-frequency features.In terms of the factors affecting the classification performance,the parameter in the SVM kernel function has the best value of 0.01,while the penalty factor C has no significant effect on the classification result;the DCT high-frequency coefficient has a significant influence on the classification result,with the DCT high-frequency coefficient.The classification accuracy rate is obviously improved.When the DCT coefficient is increased to 1,the extracted high-frequency features include DC components,and the classification accuracy rate is significantly decreased.The time used for classification is positively correlated with the DCT coefficient.The CNN-based turbulence degradation image is used.In the classification method,using the improved LeNet-5 network structure experiment,the Loss value is basically stable at around 0.66,and the classification accuracy rate is about 80%.In this paper,the DT-SVM turbulence-degraded image classification method based on DCT high-frequency features and the turbulence-degraded image classification method based on improved CNN network model are studied,and the k-value range of turbulence-degraded image classification is realized.The traditional SVM classification method affects the classification effect of turbulence-degraded images by DCT coefficients,but the overall classification accuracy is not good.The classification effect of CNN network on turbulence-degraded images is affected by the network structure,and the improved CNN network has good classification accuracy.The effectiveness of the method and the superiority of the traditional SVM classification method are verified,and the image restoration is further paved based on the classification result.