Preserving fine details in PCB x-ray imaging: a patch-based data augmentation method for improved segmentation and detection

X-ray imaging has been extensively used for printed circuit board (PCB) inspection, facilitating reverse engineering and fault detection. For effective PCB inspection, segmentation and detection of various components are important necessary steps. However, the intricate designs of PCB x-ray images and very small components pose significant challenges for accurate component segmentation, especially when using pretrained deep learning models. These models typically downsample input images during training to fit their architectural constraints, leading to the loss of critical fine details such as buried vias and narrow traces. Additionally, x-ray artifacts and occluded components exacerbate segmentation difficulties. Moreover, data scarcity is a big issue as x-ray data collection and annotation are time-consuming and costly, which further complicates the development of effective models. To address these challenges, we propose a data augmentation approach that involves slicing the original x-ray images into patches. Our approach aims to retain the subtle details and features that are often lost during image downsampling. In our experiments, we experimented with several parameters, including patch size and overlap ratio, to generate an accurate patched dataset. We implemented multiple metrics to assess the quality and correctness of the patched dataset compared to the original full-sized images. These metrics include processing time, Intersection over Union (IoU), and Structural Similarity Index (SSIM). IoU and SSIM measure the spatial overlap and structural similarity between the patched and original images, respectively. We also ensured the effectiveness of reconstructing the patches into a complete image without losing crucial details. Our proposed patch-based approach ensures that critical features and annotations are preserved, providing a reliable dataset for further downstream tasks, to fit into the training pretrained models, while addressing the challenges of downsampling of the minuscule PCB components.