Efficient Representation Of Steel Strip Surface Image Based On Texture Feature Learning

Strip steel,as an important steel product,is widely used in automobile manufacturing,bridge construction,aerospace and other pillar industries,it is of great significance to ensure the quality of its final products for the development of society and the improvement of life.The number,extent and distribution of surface damage areas are important factors in determining the quality of strip steel.The surface defects detection and classification methods based on computer vision can well realize the location and identification of strip surface defects,and discover the causes of defects,so as to effectively guarantee the quality of strip steel products.The accuracy and high speed of strip surface defect detection and classification depend mainly on the efficient representation of its surface images.This thesis studies the efficient representation of strip steel surface images and applies it to defect classification.The main research contents of this thesis are as follows:Firstly,the research methods of strip steel surface defect detection and classification in the past three decades are collated,and they are classified into several typical categories according to the principles of these methods,and summarized,analyzed and compared.It is found that the local binary pattern(LBP)method is a lightweight texture description operator that can be used for both defect detection and classification,which is important for the efficient representation of strip steel surface images.Then,a simple,fast and robust texture descriptor,which is Selectively Dominant Local Binary Patterns(SDLBP),is proposed for the efficient representation of strip surface images.An intelligent search algorithm with quantitative threshold mechanism is established to mine Dominant non-uniform Patterns(DNUPs),and two pattern code mapping schemes are developed which can be converted according to the input image quality.All the uniform patterns and DNUPs were mixed with binary encoding to further improve the accuracy and generalization ability of SDLBP for texture description,and the beneficial effect of noise suppression was obtained too.Finally,an Adaptive Region Weighting(ARW)method is designed to further enhance the Nearest Neighbor Classifier(NNC)in the feature matching stage of defect classification task.It overcomes the problem that traditional NNC lacks prior knowledge,which leads to the degradation of classification.Extensive experiments are conducted on an open texture database(Outex)and an actual strip surface defect database(Dragon),and the experimental results demonstrate that the proposed SDLBP performs well in terms of classification accuracy and time efficiency,the designed ARW method can effectively assist the basic classifier to improve the classification accuracy.Which is beneficial for the efficient representation of strip surface images.