Orange Class Identification using Neural Networks and Transfer Learning

In order to increase the sales of fruits, quality, and type of fruit play an important role. Underdeveloped countries still use traditional visual examination to inspect the type and quality of oranges as they are very similar to each other with some differences. This manual examination has a great chance of errors as human perceptions have some limitations. A recent development in the field of computer sciences, especially Machine learning and Artificial Neural networks made it possible to train the computers for the classification of oranges and other fruits. This paper presents the classification of oranges between four varieties (Kinnow, Mosambi, Shakri, and Red blood) that can be found in Pakistan in a very large amount. Dataset used consists of 400 self-captured images. Every class consists of a total of 100 images, 80% of images were used for training purposes. The proposed model was tested on 20% images. The proposed study also signifies the importance of data augmentation and Transfer Learning in the classification process. Dataset was enhanced by using augmentation transforms. By training a Convolutional Neural Network with original dataset accuracy of 80% was achieved. This accuracy of the model was improved to 83.75% when trained with augmented dataset furthermore Resnet-50 with some layers was trained with original and augmented datasets to improve the accuracy of the classifier. Accuracy of 82.5% was achieved when Resnet-50 was trained with the original dataset. This accuracy jumped to 87.5% when the augmented dataset was used.