Page 35 - ChipScale_Sep-Oct_2020-digital

defect modes and determines pass
(overkill) or reject. For an example such
as FM (particle), other defect modes of
stain, fiber, and crack can easily cause
overkill and underkill situations.
C o m b i ne d s e g me n t at i o n a n d
classification algorithms in series first
checks all rejects (by machine vision)
one more time based on customer
criteria (segmentation trained model)
and finally decides on whether the
classification should be a true reject
or pass (overkill) based on defect
classes (classification trained model)
and the customer specification. The AI
inspection progress using combined
algorithms is shown in Figure 3. As
an example, for the bump damage case
study, machine vision determines the
pass units. Reject bumps are inspected
by the segmentation algorithm, followed
Figure 3: Deep-learning inspection progress using segmentation and classification combined algorithms.
Under-reject and over-reject validations
are important steps before applying the
AI model in a real inspection scenario
li ke high-volu me manufact u r i ng
inspection. After image capturing
using vision software and collecting the
images for model learning, the training
process is completed. Next, we perform
inspection by using the AI model. For
under-reject validation, all logged images
by the review software are reviewed for
defects that were not detected. Model
learning is reinforced after labeling the
defective exported images. We then
replace the existing model in the learning
software. Finally, detection capability is
again verified by re-inspecting the under-
reject units. For over-reject validation, we
review all images logged by the review
software for overkill images. We then
reinforce model learning after adding Figure 4: Deep-learning and combined algorithms progress over an evaluation period of fifteen weeks.
the exported images without labeling.
The AI model is replaced and detection
capability is verified again. In the case
of over-reject validation, removing
similarly-labeled overkill images before
reinforcing the model learning effectively
reduces overkills.

Combined algorithms and results
O u r d e e p - l e a r n i ng a p p r o a c h
includes combined segmentation and
classification algorithms in series. Based
on customer criteria and the trained
model, segmentation determines the
defect area. Based on different defect
features (classes) and the trained model,
the classification differentiates true Figure 5: The deep-learning second-inspection process.

33
Chip Scale Review September • October • 2020 [ChipScaleReview.com] 33