Performance analysis of plant species recognition using conventional and deep learning techniques

Abstract

Real-time Plant Species Recognition under unconstrained newlineenvironments is challenging and time-consuming process. Two factors newlinesignificantly affect the efficacy of the recognition system (using leaf). They newlineinclude the complex nature of leaves and the challenges associated with computer newlinevision methods. Complex nature is due to the vast plant diversity, intra-class newlinevariations, leaf structure, leaf color (seasonal variation, aging factor), leaf type newline(simple and compound), geometry, and venation properties. The challenges newlineassociated with computer vision methods primarily relate to the image acquisition newlineprocess. The captured images exhibit illumination variations, scale changes, newlineorientation or viewpoint modifications, and different backgrounds. Furthermore, newlinein most of the reported literature, plant species recognition is performed using the newlinedatasets of foreign origin. These datasets do not include the plant images of newlineIndian origin. newlineThis research work proposes an efficient system for plant species newlinerecognition in real-time. A custom-developed dataset named Leaf-12 is formed newlineusing the Indian plant species. The dataset includes twelve plant species. The newlineimages in the dataset are captured by varying the illumination, scale changes, newlineorientation or viewpoint modification, and different backgrounds. Four newlineapproaches are tested for real-time plant species recognition. They include the newlineConventional image processing method, PReLU based Backpropagation Neural newlineNetwork (P-BPNN), Single Deep Learning (Convolutional Neural Network newline(CNN)) Architectures, and Dual Deep Learning Architectures (DDLA). The newlinemethods are evaluated using four datasets, namely, Flavia, Folio, Swedish leaf, newlineand custom-developed Leaf-12. newline