A Strategy for Real-Time Suppressing the Fluctuation of DFB Chip Microcharacter Spotting Accuracy in Optical Device Packaging

Effectively, economically and long-term stably real-time suppressing the fluctuation of microcharacter spotting accuracy (mCSA) for a submillimeter-level distributed feedback laser (DFB) chip in the complex industrial environment (CIE) is an essential but challenging task in optical device packaging. This article develops an end-to-end micro-optical character recognition net (MOCRNet) with a one-stage framework (OSF), data enhancement, and defect preprocessing. Specifically, by embedding the proposed text defect detector (TDD) and the text flip detection (TFD), the OSF has replaced the previous traditional multistage frameworks for improving long-term accuracy and efficiency. Moreover, a data enhancement technique called the environmental noise revivification method (ENRM) is applied to obtain expanded sample images with environmental noise to suppress small fluctuations by training the model; additionally, large and small fluctuations caused by defects in the character area and on characters are effectively suppressed by the background defect removal module (BDRM) and TDD, respectively. Experiments from a real-world optical device packaging line based on the proposed MOCRNet industrial intelligent microscope system reached an mCSA of better than 99.6% and accordingly decost the labor to 86% compared to 98.8% and 72% in the latest previous work, respectively. The results demonstrate the superiority of the proposed MOCRNet to the state-of-the-art models, as well as the effectiveness and accuracy, and stability for handling a great fluctuation of mCSA from submillimeter-level chips with defects and noise during the optical device packaging in the CIE.