High-speed inline defect detection for TFT-LCD array process using a novel support vector data description

摘要

TFT array process is a critical fabrication process for thin film transistor liquid crystal display (TFT-LCD) manufacturing, and defect detection plays an important role in yield improvement for this process. Due to the diversity of defect modes and their occurrence frequencies, the true distribution of the defective patterns is difficult to obtain. On the contrary, normal patterns are easy to collect and they involve only small variations in uniformity. Hence, one-class classification is an appropriate strategy for the LCD inline defect inspection. Accordingly, as a defect detector, the one-class classifier, support vector data description (SVDD), is a good candidate due to its satisfactory results in many one-class classification problems. However, SVDD has the drawback that its testing complexity is linear in the number of training patterns, which makes SVDD unable to provide a fast-enough classification speed. This is problematic because, although SVDD is accurate in defect detection task, it is difficult to implement for real-time tasks, especially when a full inspection (every LCD panel will be inspected) is required. To address this critical issue, in this paper we propose a novel SVDD, called fast SVDD (F-SVDD). The proposed F-SVDD not only inherits the merit of the traditional SVDD, which can obtain a compact description for a target set, but also can provide a much faster classification speed because its testing complexity is independent of the number of training patterns. Experimental results, carried out on real surface images of LCD panels, indicate that the F-SVDD is able to obtain a high defect detection rate of over 95%. More importantly, compared with the traditional SVDD, the proposed F-SVDD is able to accomplish the inline defect detection task with a relatively faster speed: SVDD needs to spend 30.17 min inspecting each LCD panel, while the same task can be done within 0.13 min (only 7.8 s) by F-SVDD.