Computer vision-based detection and classification of chemically ripened bananas and papayas at vendor site through deep learning AI models using real-time dataset

• Chemically ripened bananas and papaya pose significant health risks due to carcinogens and harmful substances. • Existing laboratory-based techniques and electrochemical gas sensors are cumbersome and face challenges in accurately discriminating between naturally and chemically ripened bananas and papaya on site. • The study develops an onsite device using advanced computer vision (CV) and deep learning (DL) techniques to detect chemically ripened bananas and papaya efficiently. • Among various models, CNN with Orthogonal Softmax Layers (CNN with OSL) achieved the accuracy of 91.58 % in discrimination of bananas and highest accuracy of 96.32 % in discrimination of papaya. It was implemented in real-time using Raspberry Pi, demonstrating its practical feasibility. Chemically ripened bananas and papayas, which pose health risks due to carcinogens and other harmful substances, are increasingly common worldwide. Traditional laboratory-based techniques for distinguishing naturally ripened from chemically ripened bananas and papayas face challenges due to their cumbersome processes. This study aims to address these limitations by developing an onsite device to detect chemically ripened bananas, offering a faster and more cost-effective solution. The research employs advanced computer vision (CV) and deep learning (DL) techniques, including support vector machines (SVMs), logistic regressions, recurrent neural networks (RNNs), artificial neural networks (ANNs), convolutional neural networks (CNNs), CNNs with orthogonal softmax layers (CNNs with OSLs), and regional CNNs. A custom real-time dataset of naturally and chemically ripened bananas from five different species and papaya was created for this purpose. The models were trained and tested on this dataset, with the CNN with OSL achieving the highest accuracy of 91.58 % and 96.32 % in a computational time of 2.24 s of discriminating banana and papaya. To scale up this approach, the system was implemented in real time using a Raspberry Pi board and a Pi camera, capturing live images of the fruits and processing them via trained models to detect chemically ripened fruits efficiently and accurately. This work demonstrates the potential of CV and DL techniques to combat fruit adulteration, providing a reliable and automated solution for ensuring food safety. The findings suggest that a CNN with OSL can accurately detect adulterated fruits, making it a promising tool for future developments in this field.