Page 58 - Kaleidoscope Academic Conference Proceedings 2024

2024 ITU Kaleidoscope Academic Conference

Furthermore, a comparative analysis with different
optimizers revealed promising results for the 5-layer
custom CNN, establishing it as the optimal model for this
study. Additionally, the FRCNN detection model
delivered effective outcomes by minimizing losses
associated with various traits. Therefore, we conclude
that this novel framework offers a substantial opportunity
to enhance the quality of disease analysis and detection in
agriculture.

(a) (b)
Fig. 11 (a) IoU curves of the considered classifiers (b) IoU curves
with different optimizers for 5-layer custom CNN

A comparison of various losses are done between the
considered backbone architectures of FRCNN are shown
in Fig. 12(a – d)

Fig. 12(a) Classifier loss comparison b) Objectness loss comparison

Fig. 13 Sample detected images of diseased and healthy leaves

5. CONCLUSION

Promising results are demonstrated by our proposed
approach for potato plant leaf disease detection, which is
based on a custom CNN Deep Net. High accuracy in
c) Regression box loss comparison d) Region proposed box classifying and detecting potato leaf diseases was achieved
loss comparison by leveraging deep learning techniques. For the entire work,
three classifiers were considered, namely Mobile Net, a 4-
Table – 2: Loss and IoU comparison of the 3 considered layer CNN, and a 5-layer custom CNN, which served as the
backbone architectures backbone architecture to the detection model FRCNN.
Among all the models, the best results were generated by
the 5-layer CNN, with a classification accuracy of 97.16%,
and the best IoU obtained is 0.78 under a threshold of 0.6.
An efficient and reliable way to monitor and manage
diseases in potato plants is provided by our method,
contributing to sustainable agriculture and global food
security. In future work, further improvement of the
efficiency and accuracy of our detection model is planned
by exploring advanced deep learning architectures and
optimization techniques. Additionally, the aim is to expand
Some of the detected images of the potato leaves with the our dataset to include a wider variety of potato diseases,
3 classes can be visualized from Fig. 13. The detected enabling more comprehensive detection and diagnosis
images are with the bounding box, the class name and the capabilities.
scores. The current study demonstrates significant utility
in the field of agriculture by automating the detection of
potato plant leaf diseases. Our custom 5-layer CNN
model achieved the best results in terms of accuracy and
other performance metrics, as detailed in this section.

– 14 –