Real-Time Plant Species Recognition Using Non-averaged DenseNet-169 Deep Learning Paradigm

Abstract

Real-time plant species recognition under unconstrained environment (viewpoint variation, changing background, scale variation, illumination changes etc.) is a challenging and time-consuming process. In this paper, a non-averaged DenseNet-169 (NADenseNet-169) CNN architecture is proposed and demonstrated to perform real-time plant species recognition. The architecture is evaluated on two datasets namely, Flavia (Standard) and Leaf-12 (custom created). The hyperparameters (optimizers, learning rate) are optimized to achieve higher performance metrics with lower computation time. From the experimental investigation, it is observed that Adam optimizer with a learning rate of 0.0001 (Batch size of 32) resulted in obtaining higher performance metrics. In case of Flavia dataset, an accuracy of 98.58% is obtained with a computational time of 3.53 s. For Leaf-12 dataset, an accuracy of 99% is obtained with a computational time of 4.45 s. The model trained on Leaf-12 dataset performed better in identifying the plant species under unconstrained environment.