Machine learns to tell the difference
Bananas at different stages of ripening are processed to make seventeen different kinds of food products. Selecting the banana at the right stage by visual inspection is time consuming. Automatic detection from digital images using machine learning algorithms has been attempted, but had limited success. Now researchers from Coimbatore report a break through.
Researchers from the Sri Ramakrishna Engineering College and the United Institute of Technology, Coimbatore initially had RGB and near infrared images of about 270 bananas at different stages of ripening – unripe, partially ripe, ripe and overripe. The researchers used 80% of the images in each stage and they trained the convolutional neural network algorithm to classify bananas based on colour. And they tested the accuracy by validating the learning using the remaining 20% of the images
Though training accuracy was good, validation accuracy was not as good. They needed a larger dataset to train and validate the machine learning. The team adopted a simple method to enlarge the dataset. They duplicated the images, by shifting, flipping, rotating, brightening and zooming in and out. The team thus had a bigger dataset. The expanded and original datasets helped train the convolutional neural network further.
The researchers compared the models trained on original and augmented datasets.
“The system trained with the augumented dataset had better validation accuracy”, says N. Saranya, Sri Ramakrishna Engineering College, Coimbatore.
The researchers then compared the new model with two other available machine learning models.
“The model based on the augmented dataset classified banana images better than the earlier models, especially in detecting partially unripe and over-ripe categories”, says K. Srinivasan,
The improved model for automatic detection of ripeness stage will help banana processing industries.
“It is not too difficult now to create a mobile application to classify bananas using a handheld device”, S. K. Pravin Kumar, United Institute of Technology, Coimbatore.
When the mobile application to detect the stage of ripeness in a timely and non-destructive way becomes available, it will be a boon for banana farmers also.
Journal of Ambient Intelligence and Humanized Computing,
K Sri Manjari
University College for Women, Koti, Osmania University
STEAMindiaReports: Tasty news that you can use
Science news free for Indian media to use