Automated classification of Bangladeshi flower species through image processing and transfer learning

Abstract

Flower species classification stands as a quintessential challenge in the realm of computer vision, bearing profound implications for botanical research, conservation, and ecological management. In this study, we embark on a rigorous exploration of deep learning architectures, meticulously assessing their efficacy in discerning intricate floral characteristics. Leveraging a bespoke dataset meticulously curated to encapsulate the rich diversity of floral species, we subject ResNet50, ResNet152, DenseNet121, and DenseNet201 to exhaustive scrutiny. Through this meticulous inquiry, we unearth the paramount importance of model selection in elucidating the subtle nuances inherent to floral taxonomy. Emerging from this crucible of computational inquiry, DenseNet201 emerges as the paragon of accuracy and efficiency, wielding an unparalleled accuracy of 99.61% with a minuscule loss of 0.06. Meanwhile, DenseNet121 stands as a stalwart contender, boasting a commendable accuracy of 96.56%. Beyond the realm of computational achievements, our findings resonate with broader implications for botanical research and conservation efforts. By shedding light on the transformative potential of transfer learning and DenseNet architectures in the realm of floral diversity analysis, this study not only advances the frontiers of technological innovation but also invigorates interdisciplinary collaborations poised to reshape our understanding of floral ecosystems. Through this convergence of computational prowess and botanical inquiry, we navigate towards a future where advanced technologies serve as indispensable allies in the noble pursuit of biodiversity conservation and ecological stewardship.