Image Retrieval Based On Deep Limit Learning Machine

With the arrival of the era of big data,the digital image data in daily production,life and science and technology show an exponential upward trend.Image data contains abundant content,and its expression is more vivid and accurate than text description.Therefore,facing the increasing number of digital image resources,it is an urgent problem to develop a fast and accurate retrieval system to find the real image of interest to users.In the field of image recognition,classifiers affect the performance of the whole image retrieval system.Limit learning machine is different from traditional neural network model.The algorithm has simple structure,fast learning speed,few training parameters and good generalization performance.It can effectively remedy the shortcomings of traditional feedforward neural network algorithm,such as slow learning speed,cumbersome parameter adjustment,etc.Traditional content-based image retrieval methods lack the ability of selflearning and image expression,which seriously restricts the performance of image retrieval.The depth learning model opens up a new way for image retrieval.The main contents of this paper are as follows:(1)Introducing the concept of deep PCA subspace,using image as input data of network,and using multi-layer cascaded PCA as convolution filter layer,the image is mapped to deep PCA subspace to complete unsupervised low-level feature learning and generate initial convolution feature mapping.(2)A deep multi-layer extreme learning machine algorithm is proposed.By stacking multi-layer extreme learning machine neural networks,L1 regularization is used to constrain the deep extreme learning machine,enhance generalization ability,and obtain abstract information hidden in high-dimensional data.(3)In this paper,a deep PCA subspace limit learning machine image retrieval algorithm is proposed.The feature extracted from deep PCA subspace is fed into the depth limit learning machine to obtain the sparse feature of deep subspace and realize the deep feature extraction of image.Finally,hash coding is applied to feature and fast image retrieval isrealized by encoding.In this paper,experiments are carried out on MNIST,CIFAR-10 and CALTECH 256 datasets.Through a large number of simulation experiments,compared with other deep neural network learning frameworks and retrieval algorithms,it has the advantages of simple structure and fast convergence speed.The algorithm in this paper has good results in retrieval accuracy and running time.