Research And Application On Image Feature Extraction Algorithm

Feature extraction algorithm and classification recognition technology is an important part of image processing technology,and the key technologies in the field of pattern recognition.Visual feature extraction includes color,texture and shape aspects.Outer contour information is an important characteristic parameter in shape characteristics,mainly obtained by edge detection algorithm and chain code technology.In order to make up for the inadequacy that traditional eight chain code algorithm will wrongly express outer contour information,puts forward symmetrical eight chain code algorithm and variable-direction eight chain code algorithm.Symmetrical eight chain code algorithm proposed the change point of decision rules,and used change point to select starting chain code direction,so as to accurately extract the image contour information.Variable-direction eight chain code algorithm just uses former pixel chain code direction to select the current pixel's starting chain code direction,so as to adaptively locate outer counter.Experiments show that both algorithms can improve the defect of traditional eight chain code algorithm.But to symmetrical 8 chain code algorithm only applies to images with simple boundary,which can guarantee the misjudgment rate lower than 1%,but for the images with complicated boundary,the extraction effects is the same to traditional 8 chain code.However,variable-direction eight chain code algorithm is not affected by the complexity of image contour,and the accurate rate of outer contour extraction is as high as 99%,with better robustness.Then through threshold setting and key-value pair tagging,symmetrical eight chain code algorithm and variable-direction eight chain code algorithm ensure the optimal one in images with multiple object as the one for classification and recognition.Area parameter is another important parameters of shape characteristics.Distance transformation algorithm based on chain code technology can improve area accuracy.This algorithm uses the thought of "similar contour" to express inner pixels according to the contour chain code sequence,and caculates this "similar contour" length of chain code sequence to obtain the area of the image.Experiments show that the area calculation accuracy is as high as 98%,and high robustness.In order to realize classification and recognition,feature parameters need to be suitable selected and handled by classifiers.The overall recognition rate of single classifier is low,in order to improve the classification accuracy,this paper chooses several classifiers which are suitable for target images,and establishs corresponding relationship between species and classifier,at last forms a cascading and multiple classifiers'combination structure.During the process of sample testing,through multiple classifiers' processing mechanism,each classifier just handles samples with accurate classification,and the remaining samples hands out to the next classifier,finally multiple classifiers summary classification results.Classification experiment takes phalaenopsis amabilis image as an example,combine random forest classifier,Adaboost classifier and SVM classifier to establish corresponding relationship between phalaenopsis amabilis species and classifiers,and then build a three-cascading classifiers suitable for phalaenopsis amabilis species.Results show that this one can strengthen the effect of single classifier,and the classification accuracy increases from 80%to 90.63%.