An approach to classify flue-cured tobacco leaves using deep convolutional neural networks

Convolutional Neural Network (CNN) is a Multi-Layer Perceptron Neural Network (MLP), specially designed for classification and identification of image data. MLPs are very useful but very slow for learning image features. Even for small images MLPs takes a lot of time to learn the features. On contrary, Convnets detects the features locally and propagate them to the neighboring layer so that the learning process is easier and efficient. Image reduction is a process normally used to reduce the number of learning parameters. The present paper is aimed at designing a new technique to convolve the input image, using Deep CNN algorithm and then reduce the image dimension by pooling techniques. The new technique is applied for image classification of flue-cured tobacco leaves. About 120 samples of cured tobacco leaves are taken for training the CNN and reduced the image dimensions from 1450×1680 to 256×256 RGB. Here four hidden layer CNN is considered and performed convolution and pooling on input images with sixteen, thirty two and sixty four feature kernels on first three hidden layers and fourth layer is connected to output layer. Max pooling technique is used in the model and classified them into three major classes' class-1, class-2 and class-3 with a global efficiency of 85.10% on the test set consisting about fifteen images of each group. Results from the proposed model are compared with other existing models and shown that the model performs better even with small training set.