Researches About Image Recognition Based On Supervised Pretraining NIN And Deep ELM Models

In recent years,machine learning based image recognition has been widely used,and play an substantial role in many fields.Images have a huge of information and considerable redundancy compared with other data.It is well known that the performance of machine learning methods is highly depend on the representation of the raw data.So constructing a image-recognition system requires a feature extractor that transforms the raw data such as the pixel values of an image into a suitable internal representation or feature vector from which the learning subsystem,often a classifer,could detect or classify patterns in the input.Conventional image recognition techniques require careful engineering and considerable domain expertise to extract feature which is laborious and difficult.Deep learning methods are representation learning methods that allow a machine to be fed with raw data and to automatically discover the representations needed for detection or classification.And multi-level of representation could be obtained by composing simple but non-linear modules that each transform the representation at one level which starting with the raw input into a representation at a higher,slightly more abstract level.With the composition of enough such transformations,very complex functions can be learned.For classification tasks,representations of higher layers amplify aspects of the input that are important for discrimination and suppress irrelevant variations.The deep learning methods used in image recognition has been a hot research topic and many successes have been achieved.In this paper,the MPNIN model is proposed to solve the image recognition problem.In NIN,instead of linear filters used in conventional convolutional layer,mlpconv layer was introduced to enchance model discriminability for local patches within the receptive field.Mlpconv-wise supervised pretraining algorithm,the central module of MPNIN,is similar as the layer-wise supervised pretraining algorithm.However,it has been proved that the latter one is too greedy to capture information about the target which could be avoid by the former.Futhermore,by introducing mlpconv-wise supervised pre-training to each hidden layer,the issue of so-called “vanishing” gradients has been suppressed effectively.However,the model MPNIN is time-consuming,that is,the MPNIN model which is based on BP algorithm has a slow training process.This paper proposes two new models,R2 CELM and R2ELM-LRF,which are constructed by combine the ELM algorithm with the philosophy of “stack generalization”.ELM tends to provide good generalization performance at extremely fast learning speed,and the problems of local minima and overfitting could be effectively avoided.Expecially,R2ELM-LRF model which based on local receptive fields and ELM,has the advantage of exploiting the local structure in images and learning capability without iteratively tuning inherited from ELM.