Datasets
Standard Dataset
PCB 3D inspection from 2D images
- Citation Author(s):
- Submitted by:
- Nikos Dimitriou
- Last updated:
- Thu, 06/09/2022 - 06:33
- DOI:
- 10.21227/93h5-ck64
- Data Format:
- License:
433 Views- Categories:
- Keywords:
Abstract
During Printed Circuit Board (PCB) manufacturing, it is critical to dispense the correct amount of conductive glue on the substrate LCP surface before die attachment, as the dispensing of excessive or insufficient glue may cause defects through short circuits or weak die bonding. Therefore it is critical to monitor the amount of the dispensed glue during production. This dataset contains RBG images and 3D laser scans of the glue deposits of 6 PCBs along with their corresponding volume as described in "A Deep Regression Framework Towards Laboratory Accuracy in the Shop Floor of Microelectronics."
Each PCB consists of 18 identical circuit modules ordered in two rows. There are 5 types of containers labeled **A**, **B**, **C**, **D**, and **E**, corresponding to a different glue shape and size that needs to be dispensed. Within each circuit module, there are four placeholders for each glue type thus resulting in a total number of 20 placeholders per module. On the top row of one of the PCBs dies have been attached, and hence the total number of available samples per glue type is 5 × 4 × 18 + 4 × 9 = **396**. For each type, 297 samples are kept for training and the remaining 99 for testing.
Each glue deposit is scanned twice using a 3D laser scanner to obtain ground truth volume measurements, once with a resolution of 50 micrometers and once with 20 micrometers. Both point cloud representations are used to analytically calculate the volume of the corresponding glue deposit resulting in two sets of measurements. Due to inaccuracies in the volume estimation process and measurement noise during scanning, the sparse 50 micrometer volume measurements are biased and corrupted by noise. The high-resolution 20 micrometer measurements are far more accurate and can be regarded as ground truth labels. The existence of a noisy and clean set of corresponding measurements makes this dataset suitable for the evaluation of regression algorithms under the influence of label noise.
The dataset is ordered in 5 directories, one for each glue type, whereas the samples of each type are split in two separate training and testing folders. Each glue sample is represented through an RGB image sample_{id}.jpg, a point cloud sample_{id}.ply, and the same point cloud after cropping the substate plane to isolate the glue cloud sample_{id}_cropped.ply. Paths to all samples along with the corresponding 50 micrometer and 20 micrometer measurements are available in the annotaions file annotaions.csv withn each directory.
