Research On Image Classification Method Based On Broad Learning System

Image classification is a hot topic in image processing,and many studies are devoted to improving the performance of image classification.However,how to classify images quickly and accurately is still a challenging task.In recent years,Broad Learning System(BLS)has been used to solve image classification problems because of its advantages of fast training speed and few hyperparameters.However,the classification performance of BLS is poor on complex image datasets(SVHN,CIFAR-10 and CIFAR-100).The aim of this dissertation is to improve the classification performance of BLS on complex image datasets.For the above purpose,this dissertation focuses on the Multi-Feature Broad Learning System(MFBLS)and Cascade Feature Block Broad Learning System(CFB-BLS).The dissertation firstly introduces the basic principle of BLS.Based on this,the MFBLS model is proposed by combining the traditional image classification method with BLS.In addition,this dissertation proposes CFB-BLS based on the deep learning image classification method.Finally,the dissertation evaluates MFBLS and CFB-BLS.The main research results of this dissertation include:(1)A multi-feature broad learning system.The model has two important characteristics: 1)Multi-feature extraction.The multi-feature extraction method extracts four features from the image,namely,convolutional features,K-means features,HOG features,and color features.These features reflect the global and local features of the image.2)Parallel structure.This structure has four feature blocks and one fusion block.The four feature blocks enhance each of the four features,respectively;the fusion block fuses the outputs of the four feature blocks and obtains the classification results of the model.This structure can make full use of the extracted features.(2)A cascade feature block broad learning system.The model contains three characteristics: 1)A convolution-based feature block.The feature block uses convolutional layers and "Squeeze-and-Excitation"(SE)blocks to learn feature.Multiple feature blocks are connected in a cascade way to learn more discriminative features.2)To avoid overfitting,the model introduces a Top-level Dropout layer.3)The model is trained by using Adam algorithm so that the model can work like a Convolutional Neural Network(CNN).With these features,the model combines the advantages of both BLS and CNN.Experiments show that MFBLS and CFB-BLS can significantly improve the classification performance of BLS on complex datasets.The research results in this dissertation provide new approaches for BLS-based image classification,thus promoting the application of BLS in image classification.