FPIC: A Novel Semantic Dataset for Optical PCB Assurance

Abstract

Outsourced printed circuit board (PCB) fabrication necessitates increased hardware assurance capabilities. Several assurance techniques based on automated optical inspection (AOI) have been proposed that leverage PCB images acquired using digital cameras. We review state-of-the-art AOI techniques and observe a strong, rapid trend toward machine learning (ML) solutions. These require significant amounts of labeled ground truth data, which is lacking in the publicly available PCB data space. We contribute the FICS PCB Image Collection (FPIC) dataset to address this need. Additionally, we outline new hardware security methodologies enabled by our data set.

Synthetic Data Generation for Text Spotting on Printed Circuit Board Component Images

“…PCB inspection tasks such as defect detection [25] or component classification [26] usually have more public PCB datasets that can be used for training or as a standard to compare different methods. Among the available PCB datasets, a few of them provide information for text markings on PCB components [16], [21]- [24], as described in Table. I.…”

“…I. A common limitation of [21]- [24] is the need for data cleaning before feeding the text data for model training, particularly due to unstandardized annotation format or human mistakes in the ground truth text annotations, whereas the majority of samples in [16] are defective character samples that are targeted for character-level aesthetic assessment task. Alternatively, PCB component images can be collected from actual production lines; however, data that come from the same site usually have imbalanced distributions within a character class and between character classes [17]- [20].…”

“…Also, the ground truth annotations are caseinsensitive (e.g., "intel" is annotated as "INTEL").  URL: https://zenodo.org/records/3886553 [22]  It contains 48 PCB images with over 8,000 annotated instances and over 10,000 text annotations [24].  It is a controlled-access dataset that contains 10,000 component training images and 3,000 validation images collected under different lighting conditions.…”

“… Text instances on microchip components are annotated at character level with tight bounding boxes, together with the character's orientation, legibility, and aesthetic classes (i.e., low contrast, blurry, and broken).  URL: https://github.com/chunchet-ng/ICText-AGCL FPIC [24]  It is a controlled-access dataset that contains 271 images and over 71,000 annotated instances from 97 PCBs. It has 23,218 text instances for a subset of selected images.…”

“…Nevertheless, PCB component images that come from the same site usually contain highly duplicated text markings and the number of samples in each character class may be imbalanced [17]- [20]. Additionally, accessible PCB datasets that provide information for text markings on PCB components [21]- [24] usually need data preprocessing such as removing human mistakes from the ground truth text annotations. These PCB datasets differ from scene text datasets in that the latter can be used directly for model training and as a benchmark for evaluation.…”