Deep Learning-Based Plant Recognition in Bangladesh

Abstract

Recognizing plant species from leaves is difficult due to the vast number of species, making identification challenging. Manual identification is quite a tedious and slow process. Thus, the automation of this process is essential in biology, forestry, research, education, and cities. In this paper, a new method is presented for plant identification based on the images of the leaves. We gathered a dataset by combining three datasets and obtained 13 classes and 2600 images of plant leaves of Bangladesh. To enhance the dataset, we performed the following preprocessing on the images: resizing, contrast stretching, gamma correction, background removal, edge detection, and augmentation. We evaluated four deep learning models: Among them, the four networks, including VGG19, Inception V3, Xception, and a hybrid model combining Inception V3 and Xception. These models were assessed using two preprocessing techniques: background removal with edge detection and contrast stretching with gamma correction. The Hybrid model turned out to be the best one as it provided a validation accuracy of 99.50% with contrast stretching and gamma correction. Xception followed with 96.61%, while Inception V3 achieved 97.85%, and VGG19 reached 97.23%. These findings highlight the transformative potential of deep learning models in advancing plant species recognition. The superior performance of the Hybrid model with contrast stretching and gamma correction underscores the importance of selecting appropriate preprocessing techniques and model architectures to achieve high accuracy for plant recognition in Bangladesh.