A Study On Support Vector Machinefor Image Recognition Implementation In Embedded Systems

The traditional methods of image recognition mainly use digital image processingtechnology and the principle of statistics, extract the statistic features of image and then useclassifier to recognize the image. It needs a large amount of calculation and the recognitioneffect is susceptible to the image quality. Although the artificial neural network has manyadvantages (such as parallel processing)in image recognition in recent years, butits owndefects (local minimum value, overfitting) limits its application. Support vector machine(SVM) is different from neural network and decision tree. It is based on structure riskminimization principle, considers the minimization of the empirical risk and confidencebounds. Therefore, SVM has more solid theoretical foundation, stronger generalizationabilityand excellent performance.With the development of integrated electronic technology, the development direction ofembedded system is low power consumption, small size and high performance, imagerecognition technology is widely used in various embedded system. When using SVM forimage recognition in embedded system, the disadvantage of consuming a large storageresources in the SVM training process, especially when the training samples set is big, makesthe training speed of the training stage be the bottleneck of practical applications, whichrestricts the promotion and application of the algorithm to some extent.In addition,the processof SVM training and classification involves a large number of floating-point operation,thusthe program execution time is too long and the real-time requirement of the embeddedsystemcan’t be satisfied.The embedded system used in this paper is the Samsung Corp’s S3C2440. And we useoffline training SVM parameter for image recognition in embedded system. First usingfloating-point arithmetic to train support vector machine on PC, then converting floating pointto fixed-point integer arithmetic in the embedded system, which reduces the computationcomplexity and improve the computing speed. This method brings a fast training computationspeed with certain loss of data precision.Another improved method is using integerparameters for training support vector machine. As the input vectors of support vectormachine used in this paper is the pixel values of the digital image,therefore it is not necessaryto integer processing for the training samples.It onlyneeds to convert the Lagrange parametersof standard SVM algorithm to integer values,thenputsthe Lagrange parametersas trainingprocess parameters to be optimized. This paper usesthe improved sequential minimal optimization algorithm, which improves SVM training and classification speed.