| The Unmanned Underwater Vehicle(UUV)widely uses underwater target image recognition technology,which has great research significance to improve the target detection and recognition ability of UUV.However,it is difficult to identify underwater images using classification algorithms,because underwater images have problems such as high noise,blurred details,and obtaining a large number of training samples is inconvenience.In order to improve the accuracy and stability of underwater image recognition,this thesis proposes an underwater image recognition method based on the Deep Support Vector Machine(DSVM)by researching image preprocessing,feature extraction,and classifiers.The main research contents are as follows:Firstly,aiming at the poor quality of underwater images and the inability to obtain good recognition results using basic features such as directional gradient histograms,this thesis preprocesses the image by normalization,noise removal,background removal,and enhancement,improves the images quality,and uses Scale Invariant Feature Transform(SIFT)as a feature descriptor,and the Bag of Words(BoW)model of images SIFT feature is constructed by the k-means clustering method to train the classifier.Secondly,in view of the low learning rate and overfitting caused by the feature dimension of the images is high and the number of samples is small,this thesis proposes the Feature Segmented DSVM(FS-DSVM)based on SVM,deep network learning methods,and feature selection methods.The FS-DSVM divides the high-dimensional feature vector into multiple small segments,uses multiple SVM to learn to extract deep features,and assigns a weight value to each small segment according to the importance.The multi-layer network is constructed in the above way,and multiple SVM are combined into the final classifier in the last layer.In order to achieve multiple classification,this study chooses the "One Versus One"method to extend the binary classifiers to multi-classifiers.The simulation results show that this method can obtain a good result of underwater image recognition.Thirdly,in order to solve the problem of unstable classifiers caused by the spatial distribution of the target image feature vector.This thesis uses the Stacking ensemble learning method to improve the FS-DSVM underwater image classification algorithm,this method takes the output of the FS-DSVM with different kernel functions as input to train a secondary classifier.The simulation results show that this method can improve the recognition rate and stability of the classification algorithm.Finally,this thesis proposes a weighted voting method with dual weights to solve the problem of low recognition rate caused by the difference in the ability of binary classifiers and the complicated classification situation.The method proposed in this thesis uses the confidence to calculate the weights of binary classifiers,and uses the Euclidean distance between the target and the training samples to calculate the weight of the target belonging to a certain class.In weighted voting,both weights are used to determine the target type.The simulation results show that this method can further improve the recognition rate of the multi-classification algorithm. |
