Mura Detection Based On Muti-Strategy Unsupervised Learning Algorithm And Human Visual System

Mura is a typically low-contrast defect on the TFT-LCD.The main performance of mura are no fixed shape,edge blur and low contrast area which can be perceived by human.At present,the majority of mura detection stays in artificial test phase in the industry of LCD.This artificial method has several properties which Restrict the production efficiency of production line,such as long running time,strong subjectivity and low stability.Therefore,it is important to research a rapid and stable methods to detect low-contrast mura,which conform to the human eye visual characteristic.The paper describes the research status of mura detection technology at home and abroad firstly.Second,we analyze the weak points and hard problems existing in the current mura detection algorithm.Thirdly,we do some researches on methods based on background reconstruction,and the key steps including pretreatment,background reconstruction and threshold segmentation.On the basis,we optimize the background reconstruction algorithm based on regression analysis.Otherwise,we propose background reconstruction algorithm on the basis of large sample,multi-strategy learning and threshold segmentation based on Human visual system.Experiment shows that the proposed algorithm obtain more objective background and good detection result.In the background reconstruction,most of the popular regression algorithms using the polynomial model which are only able to identify gravely abnormal points deviates from the model.It will lead to missing detection.For this problem,this paper introduces the learning probability of two-dimensional information and support vector regression(SVR).Through SVR fitting background,we can reduce the impact on background reconstruction caused by mura and the rate of missing detection.Most of the popular background reconstruction algorithm only use the test image itself,which is difficult to distinguish the inconsistencies of mura and background.It is likely to reconstruct mura as background and make the rate of missing detection higher.To solve this problem,this paper proposes leaning based algorithm,which train data on the defect free image set and obtain feature space to reconstruct background.This paper selects PCA model to train and reconstruct due to two reasons:1)The imbalance of the background itself generally characterized by global change;2)Mura generally characterized by localized changes caused by some physical reasons.The results of experiment show that this algorithm can reconstruct the inconsistent change of background effectively and keep the result not affected by mura.Since mura is a visual defects in essence,the prominent degree perceived by human eyes is the basis for judgment.Therefore,it is not ideal to detect mura by the traditional local threshold segmentation.To solve this problem,we introduce the concept of just noticeable difference(JND).Combining the characteristics of the human psychological with physiological mechanism,JND provides a quantitative basis for mura defect detection and evaluation standards from the perspective of the human eye.The innovations of this paper are as follows:1)A background reconstruction algorithm based on regression analysis is optimized.The algorithm reduces the miss detection rate and makes up for the shortcoming that the traditional algorithm cannot detect low-contrast mura.2)A background reconstruction algorithm based on large sample set and multi-strategy is proposed.The global features are extracted as well as details.Thus,we can improve the detection accuracy.3)An adaptive threshold segmentation algorithm based on human visual system is proposed.Low-contrast target region is effectively estimated.All kinds of noise and aberrance affect to the target area of image eliminate furthest.