Convolutional neural networks and transfer learning for quality inspection of different sugarcane varieties

Authors: ALENCASTRE-MIRANDA, MOISES - Massachusetts Institute Of Technology; Johnson, Richard; KREBS, HERMANO - Massachusetts Institute Of Technology

Submitted to: IEEE Transactions on Industrial Informatics
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 4/16/2020
Publication Date: 5/4/2020
Citation: Alencastre-Miranda, M., Johnson, R.M., Krebs, H.I. 2021. Convolutional neural networks and transfer learning for quality inspection of different sugarcane varieties. IEEE Transactions on Industrial Informatics. 17(2):787-794. https://doi.org/10.1109/TII.2020.2992229.
DOI: https://doi.org/10.1109/TII.2020.2992229

Interpretive Summary: In this research we employed computer vision and deep learning techniques to select and plant healthy sugarcane billets (seed), which increased the plant population and yield per acre of sugarcane. First, we used well known convolution neural network (CNN) methods to process and analyze large image datasets of healthy and damaged sugarcane billets. We then transferred the new information about these sugarcane billets to expand our results to different sugarcane varieties. We used a two-step transfer learning process to extend the information to new varieties. We compared results obtained during the transfer learning process using AlexNet, VGG-16, GoogleNet, ResNet101 architectures, as well as our own CNN architecture, to classical computer vision methods. Our goal was to determine the best approach to detect damaged and healthy billets in the shortest processing time. The best results in terms of both time and accuracy were obtained with AlexNet. For AlexNet, we compared different combinations of three sugarcane varieties in order to find the best model to identify the healthy sugarcane billets. We then reduced the number of images employed to retrain the model to determine tradeoff between time and performance. Ultimately one needs only a few dozen billets of the new variety to retrain the network. Our approach may allow a grower to rapidly classify the quality of the sugarcane billets and nearly double the number of healthy billets planted.

Technical Abstract: We employed computer vision and deep learning techniques to select and plant healthy billets, which increased the plant population and yield per hectare of sugarcane. We employed well known convolution neural network (CNN) architectures to process large image datasets and transfer learning techniques to expand the results to different sugarcane varieties. It would be very time consuming to collect and label large datasets for each sugarcane variety, for which quality inspection is needed, prior to planting. We used a two-step transfer learning process to extend the trained architecture to new varieties. We compared results obtained during transfer learning using AlexNet, VGG-16, GoogleNet, ResNet101 architectures, as well as our own CNN architecture, to classical computer vision methods. Our goal was to determine the best approach to detect damaged and healthy billets in the shortest processing time. Best results in both time and accuracy were obtained with AlexNet. For AlexNet, we compared permutations of three sugarcane varieties in order to find the best model to identify the healthy sugarcane billets. We then reduced the number of images employed to retrain the model to determine tradeoff between time and performance. Ultimately one needs only a few dozen billets of the new variety to retrain the network. Our approach led to meaningful increments in the yield per hectare ranging from 33 to 88% depending on sugarcane variety.