Page 88 - IJAMD-2-2

International Journal of AI for
Materials and Design Fruit image detection using AI

positive instances. It indicated how many actual positive the inclined plane, images were captured by a camera for
observations were correctly identified among all the true classification according to their classes, as implemented
positives, thereby showing the model’s effectiveness in by the SVM model. The camera used was an ESP32-cam
capturing relevant cases, as shown in Equation VI. 43 (Espressif Systems, China), controlled by the Arduino
Uno R3 (Microchip, United States), with the entire setup
TP
Recall = (VI) powered by a 9 V battery. The captured images were
TP + FN transmitted to the processing unit for classification using
(iv) Model F1 score the trained SVM model.
The F1 score is the harmonic mean of precision and Based on the classification results, the Arduino board
recall, providing a balanced metric that considers both (Arduino, Italy) sent commands to three servo motors
aspects. It was particularly useful for evaluating model (MG996r, TowerPro, China). These motors controlled
performance on imbalanced datasets, as it offered a a mechanical arm responsible for guiding the fruits into
better overview than individual metrics. The F1 score was different containers. Once the fruits reached the end of the
calculated using Equation VII. 44 inclined plane, the servo motors activated the mechanical
arms to direct the fruits into their respective containers.
×
2 2 (Precision Recall)
×
F1 = = (VII) Collection containers were placed at the base of the
1 1 (Precision Recall) inclined plane to collect the sorted fruits.
Precision Recall In this project, an efficient and accurate automated
In such a structured approach, the performance metrics sorting mechanism was developed. Figure 2A shows the
of different kernels were compared directly to identify the conceptual drawing of the mechanism, Figure 2B shows
most effective kernel for fruit classification. the side view, and Figure 2C shows the front view of the
mechanism, clearly depicting the system components
2.4. Model Optimization and their arrangement. Figure 3 displays the complete
It was essential to optimize a model for its performance experimental setup for the system, while Figure 4 illustrates
and accuracy toward fruit classification. The techniques the overall framework for the automated fruit sorting
applied in model development were also used to tune the process.
model, with a focus on the regularization parameter (C). This setup incorporated an automated sorting
The C value controlled the trade-off between mechanism integrated into the Arduino board, which
minimizing training error and testing error. controlled the classification results from the SVM model
and ensured efficient and accurate sorting of fruits into
(i) Impact of the value of C: A lower C value applied their respective containers. The system was specifically
stronger regularization by shrinking the coefficients designed to handle the fruits with the least damage, thereby
less aggressively, allowing a larger margin of error, maintaining the quality of the fruits during the sorting
which may result in higher misclassification rates. On process. All program codes that automated this system
the other hand, a higher C value reduced the error were displayed in Programs S1–S6 (in Supplementary File).
margin, thereby lowering misclassification rates.
(ii) Optimal C by cross-validation: Cross-validation 3. Results and discussion
was utilized to identify the optimal value of C. This
involved testing various C values using the model to 3.1. Experimental results for model detection of
determine which value achieved the best classification fruit image features
performance while avoiding overfitting. Table 1 shows the results of the fruit classification system.
The results highlight key differences in the efficacy of
2.5. Automated image sorting mechanism various features used in the image processing approach.
The automated sorting mechanism played a crucial role This was evident in the performance metrics, including
in the fruit classification system. It physically sorted the accuracy, precision, recall, and F1 score, all of which
classified fruits into their respective containers based on demonstrated that feature selection significantly impacts
the output classifications identified by the SVM model. the overall results of classification.
The system was designed to ensure efficient and accurate One of the important metrics is accuracy, which
sorting of the fruits. represents the proportion of correct predictions made
The process began by placing fruits on a slanting by the classifier. The accuracy obtained in this study was
plane that acted as a conveyor. As the fruits rolled down 90% with combined features. The result demonstrates

Volume 2 Issue 2 (2025) 82 doi: 10.36922/IJAMD025150011