Labeling Correction in Optical Inspection of Surface Mount Technology Assembly Through Data Cleaning Using StableDiffusionXL Combined with Contrastive Learning

Labeling errors were identified in the dataset collected from automated optical inspection during the Surface Mount Technology (SMT) assembly process. This paper presents a data cleaning process to correct these errors. To effectively reduce error rates in SMT production and improve product quality, it is essential to build a dataset encompassing various types of components, as printed circuit board panels contain more than one type of part. Leveraging this dataset, combined with state of the art image models, allows for precise classification of different components, significantly enhancing production quality and efficiency. The process uses contrastive learning for model training and StableDiffusionXL (SDXL) for image generation. Contrastive learning improves model performance by highlighting sample differences, while SDXL generates images when comparison samples are unavailable. And two experiments were conducted for testing. The first experiment used a prepared test dataset consisting of $\mathbf{2, 0 0 0}$ pairs of images, with 1,000 labeled as good and 1,000 labeled as defective, resulting in a balanced dataset. The results show that using our proposed method, the highest accuracy achieved with SDXL assistance was $73.45 \%$, while without SDXL, the highest accuracy reached $\mathbf{9 4 . 7 0 \%}$. The second experiment compared the difference between manual methods and our method, showing that our method achieved an overkill rate of $2.9 \%$ and a leakage rate of $1.2 \%$.