Research And Implementation On Techniques Of Incremental Testing Image Construction

In the industrial domain of achieving the quality testing on the product appearance,the machine vision based image processing algorithm being used in the production line has been played a crucial role.It is contributing to monitoring the product appearance in real time,and localizing unqualified produces.However,it is impossible to take products' attributes into consideration during the time of designing such kind of image processing algorithms,which would definitely lead to the algorithm's version updating during its actual usage.This defect would degrade the credibility and stability among each version of the algorithm.The adoption of reasonable and effective algorithm testing would help to avoid this kind of challenge.Therefore,considering the testing input,constructing a testing image dataset which has modest image number and various product appearance,has become an emergent problem.The meaning and role of constructing an image testing set for testing the algorithm of image processing have been analyzed based on the analysis of algorithm testing,feature extraction and incremental data clustering at home and abroad.By summarizing the research difficulties,the solid solutions for fulfilling the current problems are presented,mainly including the clustering algorithm for original images and increasing images,the classification fusion of both original images and increasing images,testing image selection and the building of the algorithm's platform.The proposed approach has been put into practice.In terms of image classification and management,both of the original image and incremental image clustering algorithms are proposed.The former is developed by grayscale vectorization features,LBP features and SIFT features.The image clustering can be performed based on the different combinations of those features,to provide cross-validation results for the testers,which makes for reducing the impact from the bad clustering and increase the reliability of the clustering results.The proposed incremental clustering algorithm can automatically handle the problems of image classification and selection.By calculating the distance between the image vector and the clustering center,compared with the adaptive intra-class distance threshold,the increasing images are determined to be added or repeat classifications.It is good for keeping the original clusters while updating the new clusters.Meanwhile,with the introduction of the clustering center ratio,the number of clusters can be controlled in the same level both for the original and increasing images,reducing the tedious work of multiple user interactions.In the terms of testing image selection,an image selection algorithm is designed.It guaranteed that the selected images in each cluster would have the maximized difference,and the duplicate-like images would be reduced as much as possible.In this way,the probable bad impact from wrong clusters can be effectively suppressed.By user fixed parameters of selected numbers for each cluster,weighting the similarity of selected interval,the maximized difference is achieved during the procedures of image selection.In the terms of platform building,the modules of feature extraction,feature reduction,original image clustering,incremental image clustering,image filtering are mainly developed.Each module not only complements each other for working together,but also can perform tasks independently.It is beneficial for the testers to customize the tasks based on the actual requirements and to construct more specific image testing datasets.In the meantime,the supplements of multiple user interfaces improve the convenience for constructing the dataset from enormous images.Being put into practice,the proposed methods and systems have greatly improved the algorithm testing efficiency and reduced the labor cost,with undoubted and significant application value.The main contributions of this dissertation are summarized as follows:? The construction of input data for algorithm testing is studied.Different from the traditional functional testing and procedure designing,the importance of constructing the input dataset for algorithm testing is emphasized more in this dissertation;? An incremental image test set construction oriented cluster algorithm is proposed,solving the problem of classification and management for the big data of product images;? An image selection algorithm for maximizing the intra-class difference is presented,guaranteeing the big difference among the selected images;? A system for constructing the incremental image testing dataset is built,through the rules of high cohesion and low coupling,the system has better scalability and reusability to serve the real projects.