Research On Broad Learning System With Radial Basis Map Function And Its Application Of Image Recognition

Image recognition is a hot pot in the field of computer vision.At the age of Internet,image data is increasing exponentially,and the complexity of images is also increasing.The requirements for computing power are becoming higher and higher.Traditional image recognition algorithms have achieved good experimental results driven by hardware,but still unable to meet people's needs in many scenarios.Aiming to improve rate of the image recognition and shorten the network learning time,the research on image recognition methods based on Broad Learning System has great value in theoretical and application.Most traditional deep learning image recognition methods have strict requirements on computer resources,such as large memory,extremely high processor computing power requirements,and long time for training.The deep learning network has a tendency to become more and more deeper continuously,which makes it subject to many restrictions in practical applications.On the one hand,the Radial Basis Function based Broad Learning System?RBF-BLS?is proposed to improve the recognition rate of the network.The method can be decribe as follow.Firstly,Uses radial basis neural network to replace feature nodes of Broad Learning System.Secondly,and polynomial function are used as weight matrix between hidden layer and output layer,those coefficients is training by momentum gradient descent.Thirdly,the fuzzy C-means clustering algorithm is used to solve the center and membership functions of the radial basis function.Fourthly,the image regularization term is added to the loss function to increase geometric details.The trained radial basis network is input to the enhancement layer,and then conduct reinforcement learning.Finally,the pseudo inverse is calculated by derivative operation to improve the accuracy and speed of image recognition.Experimental comparison shows that this method has better performance in Vehicle,WDBC and other data sets.On the other hand,the improved Radial Basis Function based on Broad Learning System?RBF-BLS+?is proposed to further improve the recognition rate of the network.This method splits the polynomial function weights between hidden layer and output layer of the radial basis network,and merges the x_i x_j,i???j terms with the output of the radial basis network as the input of the enhancement layer.In the way,not only can make data used for enhanced learning in enhancement layer,x_i x_j,i???j item save the sample characteristics of the data set.,but also can reduce the calculation amount of the gradient descent of the RBF network node,so as to avoid dimensional explosion,and avoid the RBF training problem arise of the RBF neural network training errors and deficiencies problems.We improve the network's ability to recognize.Experimental comparison shows that this method has better performance than RBF-BLS and other methods.In summary,our makes further research based on Broad Learning System.Changing the methods of generating feature nodes,and making some optimizations to improve the recognition rate of the network.Through the experimental comparison of multiple data sets,the significance of our work is further demonstrated.