You are on page 1of 3

Image classification using CNN

Manish Singh(2014EE10453)
November 26, 2016

1 Aim
The aim of the project is to classify the images using Deep Convolutional Network.

2 Method and model


For the purpose, I finetuned the BVLC caffenet model trained on the Imagenet Dataset.

2.1 CaffenetModel
Its is an image classifier built on the AlexNet architecture.

2.2 Training
2.2.1 Data Augmentation
Few classes had very few images and few were very similar to each other. Hence I merged the data few
similar classes. Still few classes very few images. To overcome this problem, I duplicated the images for
the classes having very less images.

Dataset given:
Backpack [ 408]
Bag [ 327]
Ballerinas [ 304]
Boots [ 511]
Brogues [ 104]
Espadrilles [ 133]
Flops [ 172]
Handbag [ 140]
Heels [ 28]
Loafers [ 641]
Pumps [ 117]
RunningShoes [ 846]
Sandals [1082]
Shoes [1873]
Sneakers [ 390]
Wallet [ 179]
Wedges [ 48]

2.2.2 DataBase used


LMDB database used to give image input to network

1
Figure 1: CaffeNet model

2
2.2.3 Finetuning
The weights were initialised from the pretrained model on Imagenet Dataset. Only last layers (Fully
Connected) was trained. This can be done by decreasing the learning rate of the model and increasing
the learning rate of FC layers.

2.2.4 Preventing overfitting


There is a lot of scope of overfitting as training data is very less. Changing the weight decay prevented
the model from overfitting. Cross validation was also used for checking the ovefitting.

3 Results
Accuracy: Top1 on the 81.12%.
Accuracy: Top5 on the 95.5%.

4 Conclusion
Even after having a highly unbalanced dataset, the finetuning of pretrained Caffenet model achieves very
high accuracy in classification task.

You might also like