SensorNet: Enhancing Fruit Contamination Detection with DL and Chemical Sensors

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Jul 1 2024

In a paper published in the journal Intelligent Systems with Applications, researchers highlighted the prevalence of harmful fruit additives by evaluating machine learning (ML) algorithms such as decision tree classifier (DTC), naïve Bayes (NB), and deep learning (DL) model named "DurbeenNet."

Additionally, they proposed a computer vision (CV) --based method integrating DL and a chemical sensor for detection. Utilizing a formaldehyde detection sensor, they collected sensor data for mango, apple, banana, and malta fruits in fresh and chemically treated states. NB achieved 82% accuracy, while their hybrid model "SensorNet," combining sensor and image data, achieved a substantially higher accuracy, surpassing "DurbeenNet." This approach enabled rapid detection of toxic substances like formaldehyde in contaminated fruits.

Related Work

Past work has extensively explored sensor-based detection of toxic substances in fruits, addressing methodologies and key findings across various studies. Addressing sensor technologies' practical deployment and integration challenges requires resilient solutions tailored for real-world applications. Integrating multiple sensor modalities with advanced computational techniques like ML and DL offers promising avenues to enhance accuracy and reliability in detecting harmful fruit substances.

Integrated Toxic Detection

The SensorNet model represents an innovative approach combining DL with sensor data integration to detect toxic substances in fruits. It features a complex architecture comprising convolutional layers for image feature extraction and fully connected layers for classification. This design enables SensorNet to process image data from fruit samples and sensor readings simultaneously, leveraging the strengths of each data type to enhance detection accuracy. SensorNet efficiently combines image-derived features with processed sensor data by integrating these components at the flatten layer, ensuring robust classification capabilities.

SensorNet's computational complexity stems from its dual processing streams: DL for image analysis and sensor data pre-processing. Convolutional layers within the DL component extract intricate image features, while pooling layers reduce spatial dimensions, optimizing feature extraction. Meanwhile, integrating sensor data involves pre-processing steps to align and combine numeric readings with image-derived features.

This integration occurs seamlessly at the flatten layer, facilitating cohesive data fusion before classification by fully connected layers. SensorNet's complex architecture imposes significant computational demands. However, it efficiently processes data using modern hardware accelerators like graphics processing units (GPUs).

SensorNet's hybrid approach offers significant advantages for real-time toxic substance detection in fruits. Leveraging ML techniques within a unified framework enhances the speed and accuracy of identifying harmful contaminants. This capability is crucial for ensuring food safety and public health, addressing concerns related to chemical adulteration in agricultural produce. As such, SensorNet exemplifies a sophisticated integration of computational methodologies tailored to meet the challenges of modern food safety monitoring, promoting confidence in fruit consumption through reliable detection mechanisms.

Hybrid Fruit Detection

The dataset was split into training and testing subsets, with the training subset utilized to train the DTC. This supervised learning algorithm constructs a tree-like structure based on features and their respective decisions to classify instances into predefined classes. The analysts employed the k-fold cross-validation method, ensuring that each subset was tested once. During training, the DTC identified patterns and relationships in the data. Performance evaluation with the testing set revealed an accuracy of 79.83%. Improvements such as parameter adjustments or pruning techniques might boost its effectiveness.

The dataset was split into training and testing subsets using k-fold cross-validation. Assuming feature independence, the NB classifier achieved an 81.92% accuracy on the testing set, with potential performance enhancements through parameter tuning or feature engineering. Notably, NB demonstrated accuracy (82%) in distinguishing fresh and chemically adulterated fruits among ML algorithms.

The previously developed "DurbeenNet" model demonstrated promising results in image analysis tasks, but subsequent evaluations revealed it was outperformed by the newly proposed "SensorNet" hybrid model. Integrating sensor data from a formaldehyde detection sensor and previously captured images, SensorNet achieved the highest accuracy at 97.03%, showcasing significant improvement and emphasizing the effectiveness of combining DL with chemical sensors for enhanced fruit contamination detection.

In the context of multi-class classification with eight classes, a confusion matrix assessed SensorNet's performance, organizing predictions and actual class labels into a grid. Graphs illustrating training and validation accuracy and loss across multiple folds helped evaluate the model's convergence, potential overfitting, and overall performance. Comparatively, SensorNet outperformed applied machine learning, deep learning, and hybrid models, achieving the highest average accuracy while maintaining a smaller parameter count.

Conclusion

To sum up, introducing the "SensorNet" model significantly advanced fruit contamination detection, achieving a remarkable accuracy of 97.03%. By integrating sensor data from a formaldehyde detection sensor with previously captured images, this hybrid approach demonstrated the efficacy of combining DL with chemical sensors for enhanced accuracy in detecting toxic substances in fruits. The findings highlighted the model's potential as a promising tool for ensuring food safety and integrity by accurately identifying chemical-mixed fruit.