Flower Classification using Deep Learning models

Dr Jayalaxmi G.N
Dr Shantala Giraddi Dr Shivanand Seeri Dr. P.S.Hiremath
School of Computer
School of Computer Department of MCA, KLE Department of MCA, KLE
science and Engineering
science and Engineering Technological University Technological University
KLE Technological
KLE Technological KLE Technological KLE Technological
University, Hubli.
University, Hubli. University, Hubli, University, Hubli.
[email protected]
[email protected] [email protected] [email protected]

Abstract— Deep learning techniques are used widespread

for image recognition and classification problems. Gradually, II. LITERATURE SURVEY
deep learning architectures have modified to comprise more Maria-Elena et al., (2008) performed a Flower Classification
layers and become more robust model for classification
on dataset with 103 classes. The method uses features
problems. In this paper, the base VGG16 model is fine-tuned
for the classification flowers into five categories, namely, Daisy, computed from local shape/texture, the boundary-shape,
Dandelion, Sunflower, Rose and Tulip flowers. The fine-tuned spatial distribution of petals, and the colour. Author
VGG16 model is trained using 3520 flower images. The model conducted study using SVM classifier with a multiple kernel
is achieved a classification accuracy of 97.67% for validation framework and achieved classification accuracy of 55.1%
set and 95.00% for testing dataset. The Kaggle dataset is used (single feature) and 72.8% (all features). Xiaoling Xia et al.,
for training, validation and testing of the proposed fine-tuned (2017) proposed method for Flower Classification using
VGG16 model. The goal of this work is to show that a proper Inception-v3 with transfer learning concept. Authors used
modified VGG16 deep model, which is, pre-trained on Oxford-I7 and Oxford-I02 flower dataset. The back
ImageNet for image classification can be used for other image
propagation neural network (BPNN) is used to train the last
data set using very small dataset without over fitting. The
VGG16 model uses mall size 3x3 filters. layer. The Classification accuracy is 95% on Oxford-I7
flower dataset and 94% on Oxford102 flower dataset is
Keywords: convolution neural network; transfer learning, achieved. Hossam M. Zawbaa et al., (2014) proposed
deep learning, Fine-tuning, VGG16. method for flower classification (eight categories) using
Support Vector Machine (SVM) and Random Forests (RF)
classifiers, which, are trained using Scale Invariant Feature
I. INTRODUCTION Transform (SIFT) and Segmentation-based Fractal Texture
Flowers are everywhere around us. They can feed insects, Analysis (SFTA) based features. Ayşe ELDEM et al.,
birds, animals and humans. They are also used as medicines (2018) proposed method for Iris Classification using a Deep
for humans and some animals. A good understanding of Neural Network with different activation Functions. The
flowers is essential to help in identifying new or rare species Relu, sigmoid and Tanh functions are used in hidden layers
when came across. This will help the medicinal industry to with epochs (100 to 400). Yong Wu, Xiao Qin et al., (2018)
improve. The system proposed in the paper can be used by conducted study using Oxford-17 and Oxford-102 flower
botanists, campers and doctors alike. This can be extended dataset. Convolution neural network models VGG-16,
as an image search solution where photo can be taken as an VGG-19, Inception-v3 and ResNet50 models are used with
input instead of text in order to get more information about transfer learning. Xie et al. (2011) proposed image
the subject and search accordingly for best matching results. classification and retrieval algorithms with LandUse-21
The importance of building automated flower recognition dataset, Indoor-67 dataset and SUN-397 dataset. Authors
method stands out in many benefits such as providing fast proposed (Online Nearest-neighbour Estimation), a unified
recognition for harvesting robots. The automated flower algorithm for both image classification and retrieval. Guru
recognition gives the people, with limited experience in et al., proposed a model for flower classification using KNN
flower species, the ability to recognize the species of a classifier. Authors created own dataset of 25 species.
flower, with the advantages of saving time and effort. Also, Segmentation is carried out using threshold based method
recently there has been a demand for flower floriculture followed by GLCM based feature extraction and Gabor
which has made it important commercial trades in filter response. Nilsback and Zisserman (2011) proposed
agriculture. This system could help in the field of method based on image features, including local
floriculture in a commercial way. shape/texture, boundary shape, colour and spatial
It is very tough mission to categorize different types of distribution of petals colour for flower classification using
flowers due to several reasons. There is lot of variation in SVM classifier with multiple kernels. Deshpande et
colors within a class of flowers and lot of similarity between al.(2019) proposed a methodology on Haar wavelet features
several classes. Even expert botanists and gardeners cannot for decease detection. Shantala et al.(2020), conducted
identify some of them accurately. Also flowers are not rigid study on fungal decease detection in maize leaves using
objects; appearance can be affected by many external factors deep learning models.
like temperature, sunlight, nutrition and humidity. There is a
need for competent automated system to identify and III. PROPOSED METHODOLOGY
recognize flower species. The proposed method for the classification of flowers is
shown in figure 1.

978-1-7281-7213-2/20/$31.00 2020
c IEEE 130

Authorized licensed use limited to: Carleton University. Downloaded on May 25,2021 at 23:25:15 UTC from IEEE Xplore. Restrictions apply.
Dataset : The Kaggle dataset is used in this study. The Fine Tuning phase, all convolution layers of the model are
dataset set contains five categories of flower images i e. trained using Stochastic Gradient Descent (SGD) optimizer
Daisy, Dandelion, Sunflower, Rose And Tulip. with learning rate (0.0001) and a momentum (0.9). The 20
Daisy : 813 images. Dandelion : 1055 images, Sunflower: epochs and batch size of 32 (number of samples propagating
734 images, Rose : 784 images, and Tulip : 984 images. through the network) .
The dataset consist of 4326 images and Images are about D. Validation: In this step, the trained model is
size 320x240 pixels. The dataset is divided into training validated or evaluated by comparing training accuracy and
dataset (80 %), validation dataset (10%) and testing dataset validation accuracy.
(10%). E. Testing: In this phase, the fine-tuned VGG16 model
The methodology consists of five steps: Data Splitting, pre- is tested by predicting the classes of given flower images
processing, training Fine-tuning VGG16, model validation retrieved from testing dataset folder.
and finally testing the model on the test dataset
Implementation: The VGG16 pre-trained model, which is
implemented using convolution layers, max pooling layers
and dense layers is used by training last three layers using
Adam optimizer with learning rate of 0.001, 20 epochs and
batch-size of 32 for flower dataset. The VGG16 consists of
13 convolution layers, 5 max pooling layers and 3 dense
layers and which sum up to 21 layers and only 16 weight
layers are considered. The small size 3x3 filters are used for
the processing the images in this model. The performance of
the pre-trained VGG16 model is enhanced by fine-tuning
VGG16 model.
The sample images used for training and validation of fine-
tuned VGG16 model is shown in the Figure 2.

Daisy flower

Dandelion flower

Figure 1. Proposed Methodology

Rose flower
A. Data Splitting: In this phase, the dataset is divided
into training dataset, validation dataset and testing dataset.
Training dataset is used to train fine-tune VGG16 model,
validation dataset is used to validate VGG16 model and
testing dataset is used to test the accuracy of the trained Sunflower
VGG16 model.
B. Pre-processing: In this phase, image features are
enhanced for further processing, and all images are resized
to 224x224 and this done using Keras pre-processing
library Image Data Generator. In this phase, flowers are
extracted from the given image. The Data augmentation is
done using Image Data Generator. Figure 2. The sample images of Kaggle dataset .
C. Training: In this phase, the top layer of the pre-
trained model was removed and replaced with a new fully
connected layer with a Softmax classifier. The Dense layer IV. CONVOLUTIONAL NEURAL NETWORKS
is the densely-connected Neural Network layer of size 1024 Deep Learning is an emerging field of research and Transfer
with the Rectified Linear Unit (relu) as the activator. In the Learning is one of its benefits. In image classification, for

International Conference on Smart Technologies in Computing, Electrical and Electronics (ICSTCEE 2020) 131

Authorized licensed use limited to: Carleton University. Downloaded on May 25,2021 at 23:25:15 UTC from IEEE Xplore. Restrictions apply.
 example, Transfer Learning makes use of features learned layer that outputs class (Five classes) probabilities. The
from one domain and used on another through feature Figure 4 shows the layers present in VGG16 model.
extraction and fine-tuning. Convolutional Neural Network Conv 1 uses 64 filters, Conv 2 uses 128 filters, and Conv 3
(also known as ConvNet) models trained on uses 256 filters. Both Conv 4 and Conv 5 use 512 filters.In
the ImageNet's million images with 1000 categories can be the proposed fine-tuned VGG16 model, all layers are trained
successfully used on other similar or dissimilar datasets, using flower dataset. The fine-tuning process adjusts the
large or small, with great success. abstract representations of the model and thus used for
The small datasets can benefit from these pre-trained flower classification.
networks because the lower layers of these pre-trained
networks already contain many generic features such as
edge and color blob detectors and only the higher layers
need to be trained on the new datasets.
Convolution Neural Network formed with the help of
different layers to perform the image classification task. The
architecture of the CNN contains the different layers as
follows:
Input Layer: Used to accept input images, and dispatched to
next layers for feature extraction.
Convolution Layer: Used to extract features from input
image using image matrix and a filter (Kernel).
ReLU (Rectified-Linear Unit): ReLU replaces all negative Figure 4: VGG16 layer definition
values in the matrix to zero, which, helps faster and
effective training. Thus, negative values are not passed to
the next layer. In transfer learning model, bottom layers in convolution
Pooling: It is used to reduce spatial size of the feature map block encode generic features whereas top layers encode
and thus reduces the parameters required to preserve specialized features based on the new dataset. Hence,
significant information. retraining of fourth and fifth convolution block in VGG16
Fully Connected Layer: Used to build feature vector of enables capture specialized features of flower dataset. The
input images and thus used for the classification of images. steps followed for fine-tuning of VGG16 model is as
Softmax Layer: It is used to normalize output of the neural follows.
network (between zero and one) and used to represent 1. Set up fine tuning on pre-trained ImageNet VGG16
network output as probability. model – train all layers of VGG16.
The first four stages are called feature extraction stages and 2. Compile the revised model using SGD optimizer
last two are called classification stages (K indicates number with a learning rate of 0.0001 and a momentum of
of classes) are shown in Figure 3. 0.9
3. Fit the revised model, log the results and the training-
time
4. Evaluate the Model
5. Save the Fine Tuning Model
6. Predict Unseen Images(Testing the model)
7. Stop

CNN Implementation: The base CNN model is implemented

using 4 Convolution layers, 4 max pooling layers and 3
dense layers with Relu and Softmax activation layer. . The
small size 3x3 filters are used for the processing the images
in this model. The model was trained with Adam optimizer,
learning rate of 0.001, The 20 epochs and a typical batch
size of 32 (number of samples propagating through the
network) were used for training this model.
Figure 3: Convolution Neural Network Layers.

In VGG16, RGB images of size 224x224 are processed by V. RESULTS AND DISCUSSIONS
five blocks of convolution layers. Each block is composed
of growing numbers of 3x3 filters. The stride = 1 is used and The proposed algorithm is implemented using fine-tuned
the convolution layer inputs are padded to preserve spatial VGG16 on Google colab. The images are in JPEG format
resolution. The blocks are joined by max-pooling layers. and in RGB color space. The Kaggle dataset is used in this
Max-pooling is applied over 22 windows with stride 2. The study for training, validation and testing of fine-tuned
five blocks of convolution layers are followed by three VGG16 model. The dataset set contains five categories of
fully-connected (FC) layers. The final layer is a soft-max flower images i e. Daisy flower dataset has 813 flower
images. Dandelion flower has a 1085 flower images,
sunflower dataset has a 734 flower images, rose flower

132 International Conference on Smart Technologies in Computing, Electrical and Electronics (ICSTCEE 2020)

Authorized licensed use limited to: Carleton University. Downloaded on May 25,2021 at 23:25:15 UTC from IEEE Xplore. Restrictions apply.
dataset has a 784 flower images, and tulip flower dataset has Processing, Bhubaneswar, India, December 2008, pp. 722–
984 flower images. The dataset consist of 4400 images and 729
Images are about size 320x240 pixels. The dataset is divided [6] Nilsback, M., Zisserman, A.: ‘A visual vocabulary for
into training dataset (80 %), validation dataset (10%) and flower classification’.Proc. IEEE Conf. Computer Vision
testing dataset (10%). and Pattern Recognition, New York, NY,June 2006, 2, pp.
The performance metric, namely accuracy is used for 1447–1454
analyzing the proposed fine-tuned VGG16 model, and is [7] Zou, J., Nagy, G.: ‘Evaluation of model-based
compared with transfer learning: base VGG16 model and interactive flower recognition’.Proc. Int. Conf. Pattern
convolution neural network (CNN). The Table 1 shows the Recognition, Cambridge, UK, August 2004, 2, pp.311–314
performance of these models. [8] Yang, M., Zhang, L., Feng, X., et al.: ‘Sparse
representation based Fisherdiscrimination dictionary
learning for image classification’, Int. J. Comput.Vis., 2014,
Validation 109, (3), pp. 209–232
Testing Accuracy
Serial No. Model Accuracy
(%)
(%) [9] Khan, F., van de Weijer, J., Vanrell, M.: ‘Modulating
1 Fine-tuned shape features by color attention for object recognition’, Int.
97.67 95.00
VGG16 J. Computer. Vision., 2012, 98, (1), pp. 49–64
VGG 16 [10] Xie, L., Wang, J., Lin, W., et al.: ‘Towards reversal-
2 (Transfer 87.78 85.45 invariant imagerepresentation’, Int. J. Comput. Vis., 2017,
learning)
3 123, (2), pp. 226–250
CNN 75.46 72.39
[11] Hsu, T., Lee, C., Chen, L.: ‘An interactive flower
image recognition system’,Multimedia Tools Appl., 2011,
Table 1. Classification accuracies of various models.
53, (1), pp. 53–73
[12] Mottos, A., Feris, R.: ‘Fusing well-crafted feature
The study conducted on classification of flowers using
descriptors for efficientfine-grained classification’. Proc.
various deep learning models namely Fine-tuned VGG16,
IEEE Int. Conf. Image Processing, Paris,France, October
Base VGG 16 and CNN. The flower dataset is from kaggle.
2014, pp. 5197–5201
The Fine-tuned VGG16 is trained from the scratch. Base
[13] Chai, Y., Rahtu, E., Lempitsky, V., et al.: ‘TriCoS: a
VGG 16 is pretrained model which is pretrained with
tri-level classdiscriminative co-segmentation method for
Imagenet dataset. CNN is convolution neural network with
image classification’. Proc. European Conf. Computer
4 Convolution layers, 4 max pooling layers and 3 dense
Vision, Florence, Italy, October 2012, I, pp. 794–807
layers with Relu and Softmax activation layer. . The small
[14] Chen, Q., Song, Z, Hua, Y., et al.: ‘Hierarchical
size 3x3 filters are used for the processing the images. The
matching with side information for image classification’.
authors obtained highest accuracy of 95% with Fine-tuned
Proc. IEEE Conf. Computer Vision and Pattern Recognition,
VGG16 which is trained with flower dataset.
Providence, RI, June 2012, pp. 3426–3433
[15] Chai, Y., Lempitsky, V., Zisserman, A.: ‘BiCoS: a Bi-
VI. CONCLUSION AND FUTURE SCOPE level co-segmentation method for image classification’.
Proc. Int. Conf. Computer Vision, Barcelona, Spain,
Authors carried out study on flower classification with
November 2011, pp. 2579–2586
various deep learning models. Highest accuracy is obtained
[16] Qi, X., Xiao, R., Li, C., et al.: ‘Pairwise rotation
with VGG fine tuned model which is trained from scatch
invariant co-occurrence local binary pattern’, IEEE Trans.
with flower dataset. To further improve the results, larger
Pattern Anal. Mach. Intell., 2014, 36, (11), pp.2199–2213
data set can be used since testing error reduces with large
[17] Hu, W., Hu, R., Xie, N., et al.: ‘Image classification
data and the model is able to generalise better from a higher
using multiscale information fusion based on saliency
amount of information.
driven nonlinear diffusion filtering’, IEEE Trans. Image
Process., 2014, 23, (4), pp. 1513–1526.
REFERENCES [19]. Deshapande, Anupama S., Shantala G. Giraddi, K. G.
Karibasappa, and Shrinivas D. Desai. "Fungal Disease
[1] Kenrick, P.: ‘Botany: the family tree flowers’, Nature, Detection in Maize Leaves Using Haar Wavelet Features."
1999, 402, (6760), pp.358–359 In Information and Communication Technology for
[2] Das, M., Manmatha, R., Riseman, E.: ‘Indexing flower Intelligent Systems, pp. 275-286. Springer, Singapore, 2019.
patent images using domain knowledge’, IEEE Intell. Syst. [20]. Giraddi, Shantala, Shrinivas Desai, and Anupama
Appl., 1999, 14, (5), pp. 24–33 Deshpande. "Deep Learning for Agricultural Plant Disease
[3] Larson, R. (Ed.): ‘Introduction to floriculture’ Detection." In ICDSMLA 2019, pp. 864-871. Springer,
(Academic Press, San Diego,CA, USA, 1992, 2nd edn.) Singapore, 2020.
[4] Chi, Z.: ‘Data management for live plant identification’,
in Feng, D.,,Siu,W.C.,, Zhang, H.J. (ED.): ‘Mutimedia
information retrieval and Management’(Springer, Berlin
Heidelberg, 2003), pp. 432–457
[5] Nilsback, M., Zisserman, A.: ‘Automated flower
classification over a largenumber of classes’. Proc. Sixth
Indian Conf. Computer Vision, Graphics &Image

International Conference on Smart Technologies in Computing, Electrical and Electronics (ICSTCEE 2020) 133

Authorized licensed use limited to: Carleton University. Downloaded on May 25,2021 at 23:25:15 UTC from IEEE Xplore. Restrictions apply.