International Journal of Artificial Intelligence and Applications (IJAIA), Vol.10, No.

4, July 2019

TRANSFER LEARNING BASED IMAGE

VISUALIZATION USING CNN
Santosh Giri1 and Basanta Joshi2
1
Department of Computer & Electronics Engineering, Kathford Int'l College of
Engineering and Management, IOE, TU, Nepal
2
Department of Electronics & Computer Engineering, Pulchowk Campus, IOE, TU,
Nepal

ABSTRACT
Image classification is a popular machine learning based applications of deep learning. Deep learning
techniques are very popular because they can be effectively used in performing operations on image data
in large-scale. In this paper CNN model was designed to better classify images. We make use of feature
extraction part of inception v3 model for feature vector calculation and retrained the classification layer
with these feature vector. By using the transfer learning mechanism the classification layer of the CNN
model was trained with 20 classes of Caltech101 image dataset and 17 classes of Oxford 17 flower image
dataset. After training, network was evaluated with testing dataset images from Oxford 17 flower dataset
and Caltech101 image dataset. The mean testing precision of the neural network architecture with
Caltech101 dataset was 98 % and with Oxford 17 Flower image dataset was 92.27 %.

KEYWORDS
Image Classification, CNN, Deep Learning, Transfer Learning.

1. INTRODUCTION
In this modern era, Computers are being powerful day by day. They have become perfect
companion with high speed computing capabilities over the time. Few decades ago it was
believed that machines are only for arithmetic operations but not for complex tasks like speech
recognition, object detection, image classification, language modeling etc. But these days,
situation is inverted. Machines are capable of doing these things more easily with very much
high accuracy. Usual algorithm consisting of finite arithmetic operations cannot provide capacity
to do such complex tasks for machine. For this, Artificial Intelligence provides lots of
techniques. Learning Algorithms are used for such purpose. Huge dataset is required for training
the model with appropriate architecture. Testing is required to evaluate whether the model is
working properly or not. Neural Network is one of AI techniques emerged long ago in 1940s but
technology at that time was not so advanced. It was up at time in 1980s with the development of
back-propagation [1]. Later it was again discarded due to slow learning and expensive
computation. Initially It was believed that only 2 to 3 hidden layers are sufficient for Neural
Network to work properly but later on it is observed that even more layers can represent high
dimensional features of the input signals [2]. Image classification has received extensive
attention since the early years of computer vision research. Classification remains main
problems in image analysis. CNNs are applicable to fields like Speech Recognition[3], text
prediction[4], handwriting generation[5]and so on.

DOI: 10.5121/ijaia.2019.10404 47
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.10, No.4, July 2019

In These days, RAM on a machine is cheap and is available in plenty. we need lots of labeled
data, time and GPU speed to train a CNN model from scratch[6]. With this limitation, it is not
feasible for training the CNN network from scratch, because it is a computationally intensive
task and it takes several days or even weeks with high speed GPU computer[7], which is not
possible with limited resource we have. Defining appropriate model for image classifications
which will produce good result in small training time and minimum CPU speed is the main task
that this paper is intended to do.

2. RELATED WORKS
In [8], Noval general K nearest neighbor classifier GKMNC[9] was used for visual
classification. Sparse representation based method[10] for learning and deriving the weights
coefficients and FISTA[11] was used for optimization. CNN-M[12], a pretrained CNN was used
for image features extractions then marginal PCA[13] is applied to reduce the dimension of the
extracted features. In[14], Alex net[15] model, a deep neural network is used to learn scene
image features. During the training phase, series of transformation such as convolution, max
pooling, etc are performed to obtained image features. Then two classifier SVM[16] classifier
and Softmax[17] classifier are trained using extracted features from the AlexNet model.

In[18], Spatial pyramid pooling was used in CNN to eliminate the fixed size input requirements.
for this new network structure SPP-net was used, which can generate a fixed length
representation regardless of image size. Standard back propagation algorithm[1] was used for
training, regardless of the input image. In[19], Kernalized version of Nave bayes Neabour[20]
was used for image classification and SVM[16] classifier was trained on Bag-Of-Features[21]
for visual classification. In[22], Extension of the HMAX[23], a four level NN has been used for
image classifications. The local filters at first level are integrated into last level complex filters
to provide a exible description of object regions. In[24], Nearest neighbor classifier[20] was
used for visual classifications. SIFT[25] descriptor to describe shape, HSV[26] values to
describe colors and MR filters to describe texture were used.

3. METHODOLOGY
3.1. Image Preprocessing

The learning method used in this experiment is supervised learning[27].In supervised

learning[27], we have to label the data for training & evaluation of the model. For training and
testing the model, Caltech101[28] image dataset and Oxford 17 flower[29] image dataset were
used. In preprocessing, all images from Caltech101[28] and Oxford 17 flower[29] are resized to
299x299X3 because to train a CNN using transfer learning[30], image input size to CNN must
be same as the input size given to original model. Then images from two standard image
dataset[28][29] were divided into training set, validation set and testing set. From caltech101[28]
dataset, amongst 70 images per class, 60 images were used for training & validation and
remaining 10 images were used for testing. For Oxford 17 flower[29] image dataset, amongt 80
images per class, 64 images were used for training & validation and remaining 16 images were
used for testing.

48
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.10, No.4, July 2019

3.2. Cnn Model Design

Fig. 1. Architecture Of Inception V3 Model [6].

The CNN model designed here is based on inception v3[6]. Fig. 1 represent the detail
architecture of Inception model. It is a 42 layer Deep, pretrained CNN model trained on
ImageNet[31], which was 1st runner up in ImageNet Large Scale Visual Recognition
Competition, with lower error rate i.e. top-1 error: 17.2% & top-5 error: 3.58 %. The inception
model has two parts; feature extraction and classification. we make use of the feature extraction
part of inception model and retrained the classification layer with Oxford 17 flower image
dataset[29] and Caltech101 image dataset[28]. To retrain the classification layer we implement
the transfer learning[30]

3.3. Transfer Learning

Inception v3[6] is a convolutional neural network model and by using GPU configured computer
it takes weeks to train from scratch[6], tensorflow[32] a machine learning framework which
provides plateform to train the classification layer with images from Caltech101[28] and Oxford
17 flower[29] using transfer learning mechanism[30]. transfer learning mechanism[30], which
keeps the weights and bias values of the feature extraction layer and removes parameters on
classification layer of inception v3[6]. First the input image of size of 299x299x3 are feded to
feature extraction layer of CNN. After that feature extraction layer calculate the feature values
for each images, feature vector are 2048 float values for each images. then classification layer of
the CNN is trained with these feature vector. The output labels in the classification layer is equal
to the number of image classes on the dataset.

Fig. 2. Transfer Learning Implementation Pipelines.

4. EVALUATION
The pretrained inception model[6] is used here for experimental purpose and the platform used
here is tensorflow[32] and the hardware plateform used here is dell latitude e6410: 2.4 GHZ
processor intel i5, 4 GB RAM. For experiment Oxford 17 flower dataset[29] and Caltech101
dataset[28] were used.

49
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.10, No.4, July 2019

4.1. Dataset

Oxford 17 flower dataset[29]: This dataset consist of 17 category flowers images with 80 images
for each class. The flowers chosen are some common flowers in Britain. The images has been
collected by Maria-Elena Nilsback and Andrew Zisserman of University of Oxford, UK.
Caltech101 image dataset[28]: This dataset has 101 classes of images. It has 40 to 800 images
per class and Collected by Fei-Fei Li, Marco Andreetto, and Marc Aurelio Ranzato in September
2003.

4.2. Evaluation Procedure

To test the model on Caltech dataset[28], 20 classes of images form Caltech101 image
dataset[28] are used. Each category consist total of 70 images. The training and validation set
consist 60 images and testing set consists of 10 images per class. And to test the model on
Oxford 17 flower dataset[29] 17 classes of flower images were used, each class consist of 80
images: 64 images per class for training & validation and 16 images per class for evaluation
were used.

To train the model transfer learning mechanism[30] is implemented on pretrained inception

model[6]. In transfer learning, weights and bias values of the feature extraction layer are kept
same as original CNN model and removes parameters on classification layer of a CNN. To train
the classification layer, training set images from two dataset[29][28] were used. The Back
propagation algorithm[1] was used to train the classification layer of the CNN model and weight
parameters of classification layer is adjusted by using cross entropy cost function by calculating
the cross entropy error between classification layer output and input feature value from feature
extraction layer.

5. RESULTS
5.1. Training

The training accuracy, validation accuracy and cross entropy graph on flower dataset[29] and
Clatech101 dataset[28] are given in Fig. 3, Fig. 4, Fig. 5 and Fig. 6 respectively. The parameters
used for retraining the model are, training steps: 4500, training interval: 1 and learning rate:
0.045. The training and validation accuracy graph in Fig. 3 and Fig. 5 represents accuracy the
CNN model get from training and validation images. And the cross entropy graph in Fig. 4 and
Fig. 6 represents the difference between the actual output and input feature value while training
the Classification layer of CNN model. In Fig. 3, 4, 5, 6 the Blue line represents accuracy &
cross entropy error variation curve on training images and Orange line represents accuracy and
cross entropy error variation curve on validation images from two dataset[29][28]. Fig. 3
represents the training & validation accuracy variation on flower image dataset[29]. Training
accuracy was 91.2% at the beginning of the training process and starts to increase, after
completion of (3/4) training steps it reached and remains to 100%. Validation accuracy was
72.13% during initiation of training and validation process and final validation accuracy was
92.7%. Fig. 4 represents the cross entropy error for training and validation respectively on
flower image[29]. The cross entropy error were 0.58 and 0.69, during initiation of training and
validation process. As the training steps increases the cross entropy error starts to decrease. At
final step of training and validation the training and validation cross entropy error were 0.02 and
0.09.

50
International Journal of Artificial Intelligence and Applications (IJAIA), Vol.10, No.4, July 2019

Fig. 3. Accuracy graph on Oxford 17 flower dataset[29].

Fig. 4. Cross entropy on Oxford 17 flower dataset[29].

Fig. 5. Accuracy graph on Caltech101 dataset[28].

51
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.10, No.4, July 2019

Fig. 6. Cross entropy graph of training on Caltech101 dataset[28].

Fig. 5 represents the training & validation accuracy variation respectively on Caltech101 image
dataset[28]. Training accuracy was 83.6% at the beginning of the training process and starts to
increase, after completion of (3/5) training steps it reached and remains to 100%. Validation
accuracy was 82.7% during initiation of training and validation process and final validation
accuracy was 94.9%. Fig. 6 represents the cross entropy error for training and validation
respectively on Caltech images[28].

The cross entropy were 1.0 and 1.3, during initiation of training and validation process. As the
training steps increases the cross entropy error starts to decrease. At final step of training and
validation process the training and validation cross entropy error were 0.28 and 0.1.

5.2 Testing and Evaluation

The classification result of the system on testing images from flower dataset[29] and Caltech
dataset[28] are represented in evaluation chart given in Fig. 7 and Fig. 8 and the results
comparisons of our system with other experiments are represented in Table 1 and Table 2.
respectively. The 'TP' in the chart is 'True Positive', which represents correctly classified images
by the system.

Fig. 7. Performance evaluation parameters on flower dataset[29].

52
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.10, No.4, July 2019

For testing the model on flower image[29], 16 images were used from each class, i.e. 272
images from 17 classes. Amongst 272 testing images, correctly classified images by the system
were 251 as shown in Fig. 7. For testing the model on caltech image dataset[28], 10 images were
used from

Fig. 8. Performance evaluation metrics on Caltech image dataset [28].

each class, i.e. 200 images from 20 classes. Amongst 200 testing images, correctly classified
images by the system were 196 as shown in Fig. 8.

Table 1. Performance evaluation of our system on flower dataset [29] with other method

S.N. Method Mean Precision

1 [24] 81.3 %
2 [33] 90.4 %
3 Our system 92.27 %

Table 2. Performance evaluation of our system on Caltech101 dataset [28] with other method

S.N. Method Mean Precision

1 [8] 59.12 %
2 [19] 75.2 %
3 [22] 76.32 %
4 [18] 80.4 %
5 Our system 98.0 %

6. CONCLUSIONS
In this paper, the classification layer of pre-trained inception v3 model was re trained
successfully by implementing transfer learning mechanism. The model yields precision of 92.27
% on 17 classes of Oxford 17 flower image dataset and 98.0 % on 20 classes of Caltech101
image dataset.

53
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.10, No.4, July 2019

REFERENCES
[1] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel,
Backpropagation applied to handwritten zip code recognition," Neural computation, vol. 1, no. 4, pp.
541-551, 1989.

[2 ] A. Krizhevsky, I. Sutskever, and G. E. Hinton, Imagenet classification with deep convolutional

neural networks," in Advances in neural information processing systems, 2012, pp. 1097-1105.

[3] D. R. Reddy, Speech recognition by machine: A review," Proceedings of the IEEE, vol. 64, no. 4,
pp. 501-531, 1976.

[4 A. Coates, B. Carpenter, C. Case, S. Satheesh, B. Suresh, T. Wang, D. J. Wu, and A. Y. Ng, Text
detection and character recognition in scene images with unsupervised feature learning,"in
Document Analysis and Recognition (ICDAR), 2011 International Conference on. IEEE, 2011, pp.
440-445.

[5] T. Varga, D. Kilchhofer, and H. Bunke, Template-based synthetic handwriting generation for the
training of recognition systems," in Proceedings of the 12th Conference of the International
Graphonomics Society, 2005, pp. 206-211.

[6] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, Rethinking the inception architecture
for computer vision," in Proceedings of the IEEE conference on computer vision and pattern
recognition, 2016, pp. 2818-2826.

[7] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A.
Rabinovich, Going deeper with convolutions," in Proceedings of the IEEE conference on computer
vision and pattern recognition, 2015, pp. 1-9.

[8] Q. Liu, A. Puthenputhussery, and C. Liu, Novel general knn classifier and general nearest mean
classifier for visual classification," in Image Processing (ICIP), 2015 IEEE International Conference
on. IEEE, 2015, pp. 1810-1814.

[9] J. M. Keller, M. R. Gray, and J. A. Givens, A fuzzy k-nearest neighbor algorithm," IEEE
transactions on systems, man, and cybernetics, no. 4, pp. 580-585, 1985.

[10] J. A. Tropp, Greed is good: Algorithmic results for sparse approximation," IEEE Transactions on
Information theory, vol. 50, no. 10, pp. 2231-2242, 2004.

[11] A. Beck and M. Teboulle, A fast iterative shrinkage-thresholding algorithm for linear inverse
problems," SIAM journal on imaging sciences, vol. 2, no. 1, pp. 183-202, 2009.

[12] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman, Return of the devil in the details: Delving
deep into convolutional nets," arXiv preprint arXiv:1405.3531, 2014.

[13] I. Jolliffe, Principal component analysis," in International encyclopedia of statistical science.

Springer, 2011, pp. 1094-1096.

[14] J. Sun, X. Cai, F. Sun, and J. Zhang, Scene image classification method based on alex-net model," in
Informative and Cybernetics for Computational Social Systems (ICCSS), 2016 3rd International
Conference on. IEEE, 2016, pp. 363-367.

[15] A. Krizhevsky, I. Sutskever, and G. E. Hinton, Imagenet classification with deep convolutional
neural networks," in Advances in neural information processing systems, 2012, pp. 1097-1105.

[16] C. Cortes and V. Vapnik, Support-vector networks," Machine learning, vol. 20, no. 3, pp. 273-297,
1995.

54
 International Journal of Artificial Intelligence and Applications (IJAIA), Vol.10, No.4, July 2019

[17] N. M. Nasrabadi, Pattern recognition and machine learning," Journal of electronic imaging, vol. 16,
no. 4, p. 049901, 2007.

[18] K. He, X. Zhang, S. Ren, and J. Sun, Spatial pyramid pooling in deep convolutional networks for
visual recognition," in European conference on computer vision.Springer, 2014, pp. 346-361.

[19] T. Tuytelaars, M. Fritz, K. Saenko, and T. Darrell, The nbnn kernel," in Computer Vision (ICCV),
2011 IEEE International Conference on. IEEE, 2011, pp. 1824-1831.

[20] K. P. Murphy et al., Naive bayes classifiers," University of British Columbia, vol. 18, 2006.

[21] Z. S. Harris, Distributional structure," Word, vol. 10, no. 2-3, pp. 146-162, 1954.

[22] C. Theriault, N. Thome, and M. Cord, Extended coding and pooling in the hmax model," IEEE
Transactions on Image Processing, vol. 22, no. 2, pp. 764-777, 2013.

[23] M. Riesenhuber and T. Poggio, Hierarchical models of object recognition in cortex," Nature
neuroscience, vol. 2, no. 11, p. 1019, 1999.

[24] M.-E. Nilsback and A. Zisserman, A visual vocabulary for flower classification," in Computer
Vision and Pattern Recognition, 2006 IEEE Computer Society Conference on, vol. 2. IEEE, 2006,
pp. 1447-1454.

[25] D. G. Lowe, Object recognition from local scale-invariant features," in Computer vision, 1999. The
proceedings of the seventh IEEE international conference on, vol. 2. Ieee, 1999, pp. 1150-1157.

[26] A. R. Smith, Color gamut transform pairs," ACM Siggraph Computer Graphics, vol. 12, no. 3, pp.
12-19, 1978.

[27] S. B. Kotsiantis, I. Zaharakis, and P. Pintelas, Supervised machine learning: A review of

classification techniques," Emerging artificial intelligence applications in computer engineering, vol.
160, pp. 3-24, 2007.

[28] L. Fei-Fei, R. Fergus, and P. Perona, Learning generative visual models from few training examples:
An incremental bayesian approach tested on 101 object categories," Computer vision and Image
understanding, vol. 106, no. 1, pp. 59-70, 2007.

[29] M.-E. Nilsback and A. Zisserman, A visual vocabulary for flower classification," in Computer
Vision and Pattern Recognition, 2006 IEEE Computer Society Conference on, vol. 2. IEEE, 2006,
pp. 1447-1454.

[30] S. J. Pan, Q. Yang et al., A survey on transfer learning," IEEE Transactions on knowledge and data
engineering, vol. 22, no. 10, pp. 1345-1359, 2010.

[31] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, ImageNet: A Large-Scale Hierarchical
Image Database," in CVPR09, 2009.

[32] M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, S. Ghemawat, G. Irving, M.
Isard et al., Tensor flow: a system for large-scale machine learning." in OSDI, vol. 16, 2016, pp.
265-283.

[33] Y. Chai, V. Lempitsky, and A. Zisserman, Bicos: A bi-level co-segmentation method for image
classification," 2011.

55