No Detail Found!

Identification of plant species through leaf vein morphometric and deep learning

Author(s)
ANEEQ ATIQUE FAST School of Computing (FSC), National University of Computer and Emerging Sciences (NUCES), FAST Lahore Campus,,Pakistan
SAIRA KARIM FAST School of Computing (FSC), National University of Computer and Emerging Sciences (NUCES), FAST Lahore Campus,,Pakistan
SAMAN SHAHID Department of Sciences & Humanities (S&H), National University of Computer and Emerging Sciences (NUCES), FAST Lahore Campus,,Pakistan
ZAREEN ALAMGIR FAST School of Computing (FSC), National University of Computer and Emerging Sciences (NUCES), FAST Lahore Campus,,Pakistan
Abstract
Taxonomy language is challenging to comprehend and automated knowledge is required to identify the plant species. The study focused on developing an improved deep neural network: Residual Neural Network-ResNet & and Densely Connected Convolution Network (DenseNet) for the plant identification with plant leaf vein architecture. There was a total of 44 species. Each species had 64 images, each of which was further divided into 52 images for the training data and 12 images for the test data. The Canny edge detection method was deployed to detect the vein architecture of the leaves. For ResNet and DenseNet, the 224 x 224 binary image was used. The size of the feature maps in 4 dense blocks was: 56 x 56, 28 x 28, 14 x 14, and 7 x 7, respectively. MalayaKew (MK) data set was used for the experiment. There was a total of 44 classes and images were divided into the training set and the test set. The training set contained 2288 images, with each class having 52 images. Test class contained 528 images, with each class having 12 images. After preprocessing these images, they were fed to various networks of ResNet and DenseNet. Two algorithms, Stochastic gradient descent (SGD) and Adam optimization, were used in each network. Through SGD, the model ResNet, had 26, 34, 50, 101, and 152 layers. The best accuracy achieved was 89.24% using 50 layers. DenseNet had 121, 169, and 201 layers. The best accuracy achieved was 94.20% using 169 layers. In Adam optimizer, the ResNet model had 26, 34, 50, 101, and 152 layers. The best accuracy achieved was 89.50% using 101 layers. DenseNet had 121, 169, and 201 layers. The best accuracy achieved was 95.72% using 169 layers. Overall, the best performance was achieved using Adam optimizer using the DenseNet model with 169 layers and came out to be 95.72%. This also surpassed the accuracy that was achieved using D-leaf architecture. The proposed deep learning (DL) methods were very accurate in identifying plants.
Publication Details
Page(s) 2195-2202
Published Journal: Pakistan Journal of Botany, Volume: 54, Issue: 6, Year: 2022

Request the full text from PASTIC

Keywords
Plant identification, Residual neural network-resnet, Densely connected convolution network-densenet
0

Citations

0

Downloads

41

Views