Page 57 - Kaleidoscope Academic Conference Proceedings 2024

Innovation and Digital Transformation for a Sustainable World

4. RESULTS AND DISCUSSION

The model framework, along with a number of
performance parameters, has been carefully examined. In
images of a mixture of healthy and diseased potato
leaves, the current study can reliably identify diseased
potato plant leaves. The present work is carried out using
a 16GB NVIDIA P100 GPU and Pytorch libraries
loaded on the Kaggle platform on kaggle notebook.
The Adam optimizer (which yielded the best results in
comparison to the other three) was used to train the
three distinct architectures which acted as the backbone
to the detection model FRCNN. For ten epochs, a batch (a) (b)
of size four is employed. The 5-layer CNN validation
accuracy was 97.16%, the 4-layer CNN accuracy was Fig. 9 (a) Confusion matrix of (b) Precision-Recall 5_layer_cnn
with the 3 classes
5_layer_cnn
73.21% and the mobilenet validation accuracy was
78.43%. The accuracy comparison graphs are displayed
in Fig. 7. The best model (5-layer CNN) then trained over
30 epochs and fed to the FRCNN model shown in Fig. 8.
Fig. 9(a) and Fig. 9(b) illustrates the creation of the
confusion matrices and precision-recall curves for the
optimal architecture. In the present work, the accuracy
comparison with the different optimizers viz. Adam,
AdamW, Sgd, Adadelta, RMS Prop has also been
executed and the best result is obtained with the Adam
optimizer as shown in Fig. 10 and reported in table – 1.
The performance parameter matrices attained potentially
superior values with precision – 97.24%, recall –
97.16% and F1 score – 97.18%. Hence, the results
obtained are promising in their values for the proposed 5-
layer custom CNN model. Fig. 10 Training and validation Accuracy comparison curves for
5- layer CNN with different optimizers

Table – 1: Accuracy comparison of 5-layer custom CNN model with
different optimizers

The FRCNN model has been deployed for the detection
Fig.7 Accuracy Curves for the Fig. 8 Training and validation purpose of the diseased and healthy potato leaves. The
3 models accuracy of 5-Layer CNN model leverages minimum losses with the best model (5-
over 30 epochs layer custom CNN) in comparison to the other models as
reported in table – 2. The best IoU (Intersection over
Union) value achieved is 0.78 under a threshold of 0.6.
The IoU comparison with the considered classifier
models as well as the IoU comparison curves with
different optimizers for 5- layer custom CNN has also
been plotted and shown in Fig. 11.

– 13 –