Research On A Feature Extraction Algorithm And Feature Reduction For Strip Surface Defect Images

Strip occupies an important position in the steel industry, It is not only widely used in automobiles, home appliances and other products needed for people’s lives,but also the important raw materials in the military,shipbuilding, aerospace and other industries. Its surface quality directly affects the final product quality and performance. The research of strip surface quality on line detection is not only of important theoreticaly innovation, but also of great value in engineering application.Defect image pattern recognition is the key to strip surface defect detection, and feature extraction is the most important step in pattern recognition. Feature extraction greatly affects the design and performance of classifier.If the features of different types vary greatly, it is easier to design the classifier which has better performance.There are many types of strip surface defects, and each has the diversity, which results that it is not easy to find the most important features,that the high separation characteristics.This makes the task of feature extraction complicated and it has become one of the most difficult tasks of structural pattern recognition system.The content and results studied in this paper are as follows:(1)The paper summarizes feature extraction method commonly used in the field of the strip surface detection,and extracted by the extracts the geometry feature、gray histogram features、wavelet transform features、GLCM features、moment features and so on by the experiment.The analysis of the advantages and disadvantages of each image feature shows that histogram features can not describe the essential characteristics of the image and moment feature presented invariant between the classes used in the fields of the strip surface defects deficiency.(2)Based on the above analysis,gives the feature extraction method based on the pulse coupled neural network (PCNN) model for mammalian visual system.The method uses the dynamic impulses characteristics of PCNN model and firing properties of neurons to extract pulse output sequence features of4-D and the ignition time signal features of2-D. It fully expresses the neighboring pixels of the gray distribution and spatial relationship. Experiments show that the extracted features have good separability,which effectively improves the recognition rate.(3)In order to more fully describe the image features,various features often need to be grouped together.But the combined features become higher dimension combined.In order to further improve the recognition efficiency and meet the real-time requirements,this paper presents a fast feature selection algorithm based on ReliefF. It uses ReliefF feature evaluation algorithm to remove irrelevant features which is not related to classification,the correlation coefficient to eliminate redundant features,and selects a group of optimum features to describe the defect information. This can quickly achieve the purpose of feature reduction and retain the distinguishing ability.It can solve real-time identification too slow caused by too much features of the defects and the recognition rate of decline caused by too complex features of the defects.In this paper, we just use the above theory and method to classify six types defect surface:edge sawtooth, welding seam, mixed material, wrinkles, pitting and abrasion. By experiments, the theory can reach average recognition rate of94.15%.