Algorithm Design Of Image Content Analysis Based On SVM

Support vector machine (SVM), which is on the basis of the theory of statisticallearning, is a new method of machine learning. It overcomes the inherent defects ofthe previous learning algorithms and provides better solutions on problems such aslimited samples problems, how to improve the generalization performance of trainingresults, how to avoid falling into local minimum and nonlinear matters in lowdimensional space.From the theory study of SVM, this paper makes use of its strong theoreticaladvantage and realizes two applications: the design of video subtitles detectionclassifier and the judgment of image object visual saliency.Semantic labels of retrieval and storage can be provided by video subtitlesbecause it expresses important information of video content, therefore, realizing videosubtitle automatical extraction is of practical application value. The algorithmproposed in this paper combines angle point density characteristics, edge strengthcharacteristics, texture characteristics and some statistical characteristics based ongray level co-occurrence matrix to distinguish between subtitle and non-subtitleregion. SVM is used to decide whether the candidate region is text or not because itnot only has better performance under the condition of small samples and twoclassification problem, but also has better generalization by avoiding setting muchthresholds used in traditional methods, such as projection.With the development of information technology, people experience the changein demand from simple text information to image information. The main body of animage mostly expresses its content and contains the largest information. Thealgorithm proposed in this paper extracts image color, edge, texture and visualsaliency map in center areas and surround areas of an image and trains a model bySVM to judge the object prominence of an image. The model can be used topre-analysis the image and give the score of its prominence degree, thus, decideswhether it is need to have the input image further processed. Experimental resultsdemonstrate that the model has higher classifier accuracy and lays the foundation ofsubsequent efficient image analysis.