Study On Appearance Defect Detection For Filters Based On Joint Information Entrop Feature Selection

Traditional artificial detection methods for appearance defects of filters can not meet theneeds of actual production, improving the detection efficiency has become a problem which isneeded to be solved.With the title “Research of appearance defect detection for filters basedon joint information entropy feature selection”, the thesis systematically analyses the methodand the key technology of appearance defect detection for filters based on joint informationentropy feature selection, which has important academic value and practical significance forapplication and development of appearance defect detection for filters.The work is funded byResearch project of Guangdong Province (01562080172294053).The paper has overviewed and analyze the technical indicators of the appearance defectdetection of filter, summarized the research progress of visual inspection of surface defects,feature selection algorithms and classification system technology, analysed the overall needsof the defect detection system, proposed the filters appearance defect detection systemframework based on joint information entropy feature selection(JIEFS).The main workincludes the following parts:①the research on functional requirements for detect Appearance defects of the filter,proposed a visual inspection solutions, build the overall vision inspection system architecture,designed the features of each part, deeply analyze the key technologies of the system.According to the characteristics of the image of filter with defect,designed a defect imageprocessing flow，proposed a hybrid feature selection model to select defect feature of filterand trained classifier.②Against the existing feature selection algorithm based on information theoryconsidering the correlation between only a single characteristics and categories, JIEFSalgorithm is proposed to explore the correlation between multiple defect characteristics andcategories. Verified the performance of the algorithm JIEFS through10data sets, andcomparative analysze with three typical feature selection algorithm BIF,MIFS, FCBF whichbased on information theory;③Discussed the mechanism of multiple classification system based on support vectormachine (SVM),and proposed a DAG-SVM multi-classification scheme which applies to theappearance of defects for filter. Research the kernel function parameters optimizationselection of SVM, finally used cross-validation method to select SVM kernel functionparameters; ④Based on the analysis of platform architecture and hardware structure of opticalsystems, selected the CCD camera and light, tested the feasibility of t defect detection systemfor filter.The results show that classification performance of JIEFS algorithm has increased by2.53%,3.05%,1.84%than BIF, MIFS,FCBFon the10data sets, and the algorithm is suitablefor large samples, quickly selecting the most relevant features.The designed detection methodfor defect detection of filters based JIEFS can correctly identify four types of filters defects,including point, mark, scratches, chipping, defect recognition rate is100%correct.