Plant identification has several applications in the area of agriculture, ethnopharmacology, horticulture. There are existing systems that identify the plant using leaf images. Because leaves are one of a plant's easily recognisable characteristics, they are frequently used for identifying by prediction based Machine learning algorithms or by performing trait segmentation by extracting information from plant’s leaf. As the structure and features of the leaf may be affected by various stages of leaves, different colors during stages, torn leaves etc. So instead using a leaf image, entire plant image is far convincing approach. Dataset is prepared by images of some monocots which are collected manually. As the count of collected images are less in number, Image Augmentation is used to generate new images with different perspectives. A neural network is used for feature extraction and to identify the plant. As the image contains a lot of features needed to be extracted, we need a neural network which is suitable. Residual Networks (ResNet) is a worthy approach as it is a way to handle the vanishing gradient problem in very deep CNNs. The proposed system is developed using ResNet-50 where the pre-trained weights are imported and the model is trained over the prepared dataset, where the model recognizes the respective species of monocot by capturing the whole image and achieved an accuracy of 99.3% using ResNet-50.
Keywords- Leaf, Trait segmentation, Monocot, Image Augmentation, Feature Extraction, Neural Network, Residual Networks, Vanishing gradient problem, Deep CNNs, ResNet-50.