Automated Classification Of LCD Micro-defect Based On Feature Combination

Thin-film transistor liquid crystal display(hereinafter abbreviated as TFT-LCD),because of its excellent color performance, ultra-low power consumption, and the characteristics of light, have been successful in the field of display devices. In LCD mass production process, precise control of the quality of the work is to ensure uninterrupted production line, which depends on the real-time quality control analysis of the semi-finished products each stage of the production line.Process of TFT-LCD process are most likely to produce defects in array process,which contains complex steps. An integrated screen often contains hundreds of millions of transistors and are arranged in a periodic array. Depending on the production process, it is used to be superimposed 4-7 layer process and these layers are superimposed on a sheet of glass. Defects in this process, if not promptly found and solved,result in irreparable damage. Process array consists of three basic operations to achieve:namely, coating, lithography and etching. The three step process may produce defects on the different types. These defects according to their origin can be divided into many categories. In the production of LCD production, the timely discovery and identify their category, for discovery of the defect generation process, improving the generation process, improve production quality are essential.In the past, test, manual inspection has been playing an important role, after the Array manufacturing process. It requires a lot of testing people to identify possible defects with the naked eye. This work is not only extremely heavy, and because of differences in people’s judgment, there will be a lot of false positives.In this article, we propose how machine learning and computer vision solutions can help the detection of semi-finished LCD products. We first describe different characteristics of LCD defects, and give several reasonable implementation of e?-cient descriptors. Then we propose a more suitable coding process which is suitable for defect image, and an advanced feature fusion method to make full use of all features. Finally, we give the technical details of the algorithm in the project implementation,including the optimization of parallelism, e?cient linear calculations and so on. These technologies enables our algorithm able to handle a large quantity of data.