The Research And Application Of Melting Reduction Auto-Encoder Algorithms

Objective: In recent years,deep learning based on neural networks has achieved rapid development in many fields such as computer vision and natural language processing and has now become the core technology of artificial intelligence.Auto-Encoders(AE),as one of the representatives of unsupervised algorithms in deep learning,has the advantages of simple network structure and neural network support,and can automatically extract features from a large amount of unlabeled data.At present,various versions of autoencoders based on traditional autoencoders have made some progress in different aspects,but there are still common problems such as weak feature expression ability,excessive decoding loss,low model learning efficiency and weak robustness.This paper is devoted to the innovative research of auto-encoder algorithms.Based on the inductive analysis of various existing auto-encoder algorithms,the advantages and disadvantages and future development directions of existing auto-encoder algorithms are clarified.At the same time,an autoencoder algorithm that can be applied to different types of image feature extraction,has better feature expression ability and fewer model parameters is proposed,and the proposed algorithm is verified by experiments on different types of image datasets.Methods and results: Firstly,a Melting reduction Auto-Encoders(MRAE)algorithm is proposed for the existing auto-encoder feature extraction and reconstruction loss optimization;Secondly,a De-Melting reduction AutoEncoder(DMRAE)algorithm is proposed to solve the problems of single feature extraction and fusion strategy and difficulty in optimizing local information.(1)Existing auto-encoders generally have problems such as insufficient feature extraction,large number of model parameters,and low model learning efficiency.Aiming at these problems,this study proposes a reduced autoencoder algorithm.First,an algorithm framework of "melting and reducing network structure" is designed,which realizes the effective fusion of local features and global features through skip connections in the encoding stage,and reduces the feature decoding loss by reducing the number of decoding layers in the decoding stage.It can enrich the features extracted by the model and reduce the overall parameters of the model.Secondly,a joint reconstruction loss function is designed.By jointly optimizing the reconstruction loss between the symmetric feature layers and the reconstruction loss between the signals,a new algorithm optimization idea is formed,which can improve the learning efficiency of the model at the same time.Effectively avoid model prematurity.Experiments show that the features extracted by the ablation autoencoder on the lung CT image dataset can be used in support vector machines(SVM),k-means clustering(K-Means)and classification and Classification and Regression Tree(CART)and other classification algorithms,the accuracy of pneumonia screening is above 97%;the features extracted on the Cats vs.Dogs data set have a classification accuracy of more than 90% in the fully connected classifier.,in different classification algorithms,the classification accuracy of features extracted by similar algorithms is significantly better,which not only shows that the subtractive autoencoder has strong feature extraction ability,but also shows that it has better generalization.At the same time,after the convolutional autoencoder is optimized by using the reduced network structure and joint reconstruction loss function respectively,the classification accuracy of the extracted features is also significantly higher than the original algorithm,which indicates that the reduced network structure and joint reconstruction The loss function significantly improves the model performance.(2)In order to further improve the adaptability of the model to features at different scales and the effectiveness of feature fusion,this study proposes a decomposed subtractive autoencoder based on the subtractive autoencoder.First,the model proposes a joint ablation network structure in the network structure part,designs a parallel dualbranch feature extraction framework in the coding stage,and adopts serial and parallel fusion strategies within and between branches to enhance the features extract performance of the model.Secondly,in terms of model optimization,a decomposition and reconstruction loss function is proposed,a target partition optimization strategy is designed,the joint optimization scheme of the symmetric feature layer and the reconstruction loss between signals is improved,and the model learning efficiency is improved.Finally,on different types of images,such as breast cancer axillary lymph node ultrasound images,lung CT images,and natural weather images,the performance tests of the decomposition and subtraction autoencoder,the joint ablation network structure and the decomposition and reconstruction loss function are carried out.The results show that the classification accuracy of the features extracted by the decomposition and subtraction auto-encoder is significantly better than that of the similar algorithms on three classifiers such as fully connected,SVM and CART,and the Peak Signal-to-Noise Ratio(PSNR)and Structural Similarity(SSIM)of the decoded reconstructed images are also better outperforms other algorithms.Conclusion: Based on the systematic analysis of the existing autoencoder algorithms,in order to solve the common problems of the existing auto-encoder algorithms,this paper proposes a subtractive auto-encoder and a decomposition-reduced auto-encoder algorithm successively.Among them,the construction of the reduced auto-encoder provides a solution and research basis for solving the problems of existing auto-encoder algorithms.The decomposition and subtraction autoencoder further deepens the research on the autoencoder model structure and reconstruction loss function,and forms an effective technical method and processing tool that can perform unsupervised feature extraction on different image data.The decomposition and subtraction auto-encoder has a certain role in promoting the intelligent processing of image data.