A Cost-Effective Framework to Predict the Ripeness of any Fruit Based on Color Space

Fresh or ripe fruits are distinguished from rotten ones by their appearance which includes color, size, texture among other things. However, the factor that carries the most weight is color. Farmers and industrial markets spend hours on labor used to handpick particular fruits from a wide variety based on their outlook. However, this task is tedious and prone to human errors. Current models that exist for automating this process are computationally expensive. Hence, this paper proposes a cost-effective framework that can be implemented on low powered embedded systems. The proposed system achieves this by first isolating the subject of an image from the background using an image segmentation method called GrabCut. Afterwards, it extracts the top three most dominant colors of a fruit in the HSV color space via k-means clustering. Finally, a k-nearest neighbor classifier is first trained on the labelled data generated beforehand. Then, it is used to predict the ripeness level of subsequent images. All these steps take place in real time. To validate the proposed model, a publicly available image dataset of bananas with 4 different ripening levels was used, wherein it obtained an accuracy score of 85.0%, processing each image in well under a second in a Raspberry Pi 3 Model B.