An Image-Set of Coffee Berries for CNN Classification

Selective picking and post-harvest sorting of coffee berries is a labour intensive exercise. If applied correctly, it can lead to higher prices and increased returns for farmers. The increased memory and processing capacity of embedded and mobile devices, coupled with the development of lightweight neural networks has created the possibility of the application of CNNs in the automation of these processes. Such a development would require development of a specialised CNN object detector, which would in turn require image data for training and testing. This work develops a database of 5000 RGB images of coffee berries in six classes. The database has been applied to train and test the CNN GoogLeNet and the Lightweight CNN ShuffleNet. The trained networks achieved more than 94% mean precision in classification of the test set in both cases. The training time for the lightweight ShuffleNet was about one third of the training time for Go ogLeNet. The time taken for the classification was not observed to be significantly different between the two networks. ShuffleNet was found to have 2.7% lower precision and recall than GoogLeNet.