Algorithm Research For Image Classification Based On Wide-depth Fusion Learning Networks

Image classification,as an important branch of machine vision,has been widely applied and achieved great success in health care,industry,agriculture and other fields.In the past decade,with the growth of data volume and the development of computer hardware technology,deep neural networks for their automatic feature extraction ability have attracted the interest of many researchers and become the mainstream algorithms of image classification.However,most of deep neural networks suffer from the timeconsuming training process because of complicated structures.To achieve this mission,a large amount of data and high computing hardware(such as GPU or TPU)have been involved.Although broad learning networks represented by broad learning system have a short training process under ordinary computing hardware CPUS due to simple structures and easy to be expanded,the research of them on image classification is still in its infancy,and the classification accuracy needs to be improved.In view of this,how to effectively and efficiently classify images with lower hardware computing costs remains a challenging but practical significance and theoretical value task.This dissertation proposes wide-depth fusion network structures and algorithms with good classification property,according to the effects of width and depth of networks on image classification performance,and combining theories such as sparse encoding,ridge regression,and random singular value decomposition(RSVD),etc.The main research content and contributions of this dissertation are as follows:1.To improve learning efficiency for large data sets with simple background,a broad-based ensemble system is proposed.It is established in the form of flat network,where the flatted input consists of the concatenation of compact and sparse features obtained by Lasso sparse encoding,and the output-layer weights can be solved by Ridge Regression Learning Algorithms.Broad incremental learning algorithms are also developed for the dynamic updating for feature nodes and feature groups when the system deems to be necessary.Simulation results demonstrate the effectiveness of the proposed system.Also,it verifies the positive effects of sparse encoding,expansion in wide sense,and a certain type of excitation function on classification performance.2.To enhance feature extraction,a 4 incremental learning method of broad learning system is developed.The 4 increment includes additional feature nodes,enhancement nodes mapped from additional nodes,all feature nodes and the original enhancement nodes.A large number of comparative experiments verify that double mapping of enhancement nodes and the feature diversity extracted can effectively improve classification accuracy.3.In order to enhance feature extraction,two flat wide-depth fusion networks are proposed.Multi-feature incremental algorithms,namely 3 feature incremental method,4feature incremental method,and 5 feature incremental method,are also developed for the updating of the two networks.A series of contrastive simulation on PC demonstrate that increasing network depth and feature diversity can enhance feature extraction ability to improve image classification accuracy.4.In order to reduce redundant features,a method using random singular value decomposition to simplify structures is proposed for broad learning system and wide-depth fusion networks.It could reduce dimensionality of feature matrix and flatted inputs.Besides,power iteration random singular value decomposition of broad learning system is proposed for noisy image data sets.Numerous simulations results demonstrate the proposed methods can effectively simplify network structure and remove redundant features to improve image classification accuracy.Finally,the research content and future research work of this dissertation are summarized and prospected.