At present, manual processing is done for tobacco leaves harvesting and segregation. Real time detection and classification of tobacco leaves is one of the biggest challenges in tobacco cultivation. Fast and accurate classification system of uncured tobacco leaves, helps farmers during curing process with less computational resources. In this paper, a fast reliable and accurate pipeline for real time tobacco leaves classification using deep neural networks is presented. In the proposed work, we have used our own captured Indian soil produced tobacco leaves image data set and classified into several classes based on the curing process. The model is trained and compared using various deep learning architectures based on their top-1 and top-5 accuracy. The proposed model is trained with 720 green tobacco leaves, which achieved 92% of top-1 accuracy after implementing various data augmentation techniques. CNN is one of the best performance models and is used extensively for image classification purposes. One of the major characteristics of CNN is its self extracting feature property which hugely helps in segmentation and classification problems. In our proposed system, we compare some of the well known pre-defined CNN architectures such as ImageNet, GoogleLeNet, VGGNet, AlexNet, Inception V3 and V4, ResNet, DenseNet, ConveNet etc trained and tested on our dataset. Ourmodel is further run on the Raspberry Pi 3 for real time classification.

Classification, Features, Tobacco leaf, Grading, Deep learning, Convolutional Neural Networks (CNN).