IJCA Vol 4 i1 2025 webmag - Flipbook - Page 30
30 The International Journal of Conformity Assessment
A value of 100% means that all
component IDs (homologation
labels) are correctly detected
and no identification errors have
occurred (100% true positives).
metal labels, and plastic labels.
In comparison, the GPT-4 Vision
system achieved an approximate
accuracy of 100% for almost
all component classes. GPT-4
The OCR analysis yielded different Vision correctly recognized the
results. Under the PoC conditions,
information for metal labels,
the OCR models Tesseract
plastic labels, elastomers with
and EasyOCR had limitations in
color printing, and color printing on
accurately detecting component
glass, achieving 100% accuracy.
IDs. These models had the highest For the remaining component
number of false positives. Only
classes such as metal molding,
for Component Class 1 labels
metal engravings, transparent
and Component Class 2 labels
plastic, black plastic, and color
did all OCR models achieve 100%.
printing on molding, over 80%
However, for many images, the
of the component photos were
OCR models were unable to provide correctly recognized.
accurate results, often recognizing The next step is to evaluate
only individual letters or digits.
the different models under real
Identification errors included
conditions.
confusion between characters
Under the real conditions at the
such as "A" and "H," as well as
assembly line, a slight decrease in
confusion between numbers and
performance of the OCR models
"o" with "0."
and the GPT-4 Vision system was
Under the PoC conditions, the
observed. In particular, for metallic
Tesseract and EasyOCR models
components (molding/engraving)
had limitations in recognizing metal and thermoset/thermoplastic
engravings, transparent and black
components (transparent/black),
plastic embossing, and elastomers an average decrease in results of 2
with color printing or molding. In
percentage points was noted. The
particular, the embossing was a
detection of labels, elastomers, and
challenge to detect. The PaddleOCR glass remained unchanged.
model consistently delivered a
Possible causes for these
higher percentage of successfully
deviations could be the different
detected component IDs but also
lighting conditions in this area.
did not achieve 100% accuracy
These conditions could lead to
except for color printing on glass,
overexposure and reflections,
Figure 13: RFID - Results ideal conditions
thereby affecting the detection and
interpretation by the technology.
Another explanation could be that
the specific lighting conditions
highlight or attenuate certain
features of the components,
leading to misinterpretations
(Guoping Li et al., 2006).
It is crucial that the model used
does not have identification
deviations, meaning it should
avoid incorrect interpretations
of numbers and discrepancies
in spacing. To achieve the
required results according
to the requirements profile,
improvements in photo capture or
image processing are necessary.
5.2.2 Results for TransmitterReceiver Systems
Figure 13 presents the evaluation
of the transmitter-receiver
systems. The results are tested
under ideal conditions (metal box,
as described in the framework
conditions in Section 5.1.2). The
following statements always
refer only to the conditions and
objectives of the PoC and do not
represent an assessment of the
tags themselves.
On the ordinate axis of the graph,
the minimum activation power in
dBm is depicted, which represents
the smallest required power level to
activate an RFID system. The value
in dBm indicates the strength of the
signal needed to activate the RFID
chip and enable data transmission
(Silva et al., 2018). On the abscissa
axis, the component classes from
1 to 4 are compared according to
the defined component classes
(metal, thermoset/thermoplastic,
elastomer, glass) as discussed in
Section 4.2. Each component class
shows the results of the minimum
activation power in dBm for the
corresponding tags.
