An Image Classification Algorithm Based On Block KPCA And Extreme Learning Machine

Computer vision is a science that how to use the machine to understand theworld. With the development of network technology and the popularity of mobileintelligent terminal equipment, visual information appears in large numbers. How touse computer vision technology to realize the analysis of visual information, and theneffectively carry on the organization, expression, management and retrieval.Becoming an urgent problem to be solved in the current scientific research andindustry. Among them, the most classic image segmentation technology has beenmore and more attention, image segmentation is the basis of image analysis,understanding and recognition, segmentation results will determine the performanceof the subsequent image analysis and recognition. Currently, the feedforward neuralnetwork based on traditional gradient training algorithm of the principle exists manyproblems in dealing with image segmentation, such as need multiple iterations, easyto fall into local minimum, need clear performance indicators and the learningdetermining, which becomes the main bottleneck restricting its development. Huanget.al put forward a new kind of single hidden layer feedforward neuralnetworks—Extreme Learning Machine (ELM). It is increasingly favored by manyresearch scholars due to its simple structure, fast learning and good generalizationperformance advantages. This dissertation will introduce extreme learning machine(ELM) method into the image segmentation problem, putting forward a kind of imagesegmentation algorithm based on extreme learning machine.The thesis is mainly divided into three parts: the first part introduces the researchbackground and significance of the image and research status; The second partdescribes the KPCA theory, we use the block KPCA algorithm based on theshortcomings of the KPCA while use ant colony algorithm to optimize the parametersof KPCA for image feature extraction,and finally use extreme learning machineclassifier for image classification; The third part describes compressive samplingtheory and two types of compressive sampling problems, we use compressivesampling techniques to optimize compression the input parameters and limits of thehidden nodes of the ELM. Finally, the simulation results show the feasibility and effectiveness of the optimized network. This paper innovation points are as follows:1. As a global method KPCA，when a sample is considered to extract features ofthe overall information of the specimen rather than the local characteristic.When thelocal feature is very important, for example different lighting situations,useing KPCAfeature extraction may lose some very important local information.Based on thisfoundation,this paper uses block KPCA (MKPCA) Strategy.MKPCA is better thanKPCA.The basic idea is to divide sample into several blocks, then use KPCA featureextraction for each block.2. In the KPCA algorithm,we give a kernel function and the kernel function mayexist one or several nuclear parameters, a different parameter values will affect thefeature extraction results and the final results of the pros and cons.Therefore, the mostimportant issue of the kernel method is the nuclear option parameters of the problem.This paper uses particle swarm optimization algorithm to optimize the KPCA nuclearparameters.3. This paper introduced compressive sampling theory, using compressionsampling techniques to optimize the learning network. The idea of the algorithm is tocompress the sample the limits of the learning network and come ture the purpose ofnetwork parameters optimization;At the same time, we use the proposed algorithm tosparse processing the output right of the network and achieve the optimization ofnetwork hidden node. Through the above steps, the limit can be a good learningnetwork structure optimization, reducing the computational complexity limitslearning.